Librdkafka Changelog
This page is a copy of the releases of librdkafka
.
2.5.3 (2024-09-02)
librdkafka v2.5.3 is a feature release.
- Fix an assert being triggered during push telemetry call when no metrics matched on the client side. (#4826)
Fixes
Telemetry fixes
- Issue: #4833 Fix a regression introduced with KIP-714 support in which an assert is triggered during PushTelemetry call. This happens when no metric is matched on the client side among those requested by broker subscription. Happening since 2.5.0 (#4826).
Checksums
Release asset checksums:
* v2.5.3.zip SHA256 5b058006fcd403bc23fc1fcc14fe985641203f342c5715794af51023bcd047f9
* v2.5.3.tar.gz SHA256 eaa1213fdddf9c43e28834d9a832d9dd732377d35121e42f875966305f52b8ff
Note: there were no v2.5.1 and v2.5.2 librdkafka releases
2.5.0 (2024-07-10)
[!WARNING] This version has introduced a regression in which an assert is triggered during PushTelemetry call. This happens when no metric is matched on the client side among those requested by broker subscription.
You won't face any problem if: * Broker doesn't support KIP-714. * KIP-714 feature is disabled on the broker side. * KIP-714 feature is disabled on the client side. This is enabled by default. Set configuration
enable.metrics.push
tofalse
. * If KIP-714 is enabled on the broker side and there is no subscription configured there. * If KIP-714 is enabled on the broker side with subscriptions that match the KIP-714 metrics defined on the client.Having said this, we strongly recommend using
v2.5.3
and above to not face this regression at all.
librdkafka v2.5.0 is a feature release.
- KIP-951 Leader discovery optimisations for the client (#4756, #4767).
- Fix segfault when using long client id because of erased segment when using flexver. (#4689)
- Fix for an idempotent producer error, with a message batch not reconstructed identically when retried (#4750)
- Removed support for CentOS 6 and CentOS 7 (#4775).
- KIP-714 Client metrics and observability (#4721).
Upgrade considerations
- CentOS 6 and CentOS 7 support was removed as they reached EOL and security patches aren't publicly available anymore. ABI compatibility from CentOS 8 on is maintained through pypa/manylinux, AlmaLinux based. See also Confluent supported OSs page (#4775).
Enhancements
- Update bundled lz4 (used when
./configure --disable-lz4-ext
) to v1.9.4, which contains bugfixes and performance improvements (#4726). - KIP-951 With this KIP leader updates are received through Produce and Fetch responses in case of errors corresponding to leader changes and a partition migration happens before refreshing the metadata cache (#4756, #4767).
Fixes
General fixes
- Issues: confluentinc/confluent-kafka-dotnet#2084 Fix segfault when a segment is erased and more data is written to the buffer. Happens since 1.x when a portion of the buffer (segment) is erased for flexver or compression. More likely to happen since 2.1.0, because of the upgrades to flexver, with certain string sizes like a long client id (#4689).
Idempotent producer fixes
- Issues: #4736 Fix for an idempotent producer error, with a message batch not reconstructed identically when retried. Caused the error message "Local: Inconsistent state: Unable to reconstruct MessageSet". Happening on large batches. Solved by using the same backoff baseline for all messages in the batch. Happens since 2.2.0 (#4750).
Checksums
Release asset checksums:
* v2.5.0.zip SHA256 644c1b7425e2241ee056cf8a469c84d69c7f6a88559491c0813a6cdeb5563206
* v2.5.0.tar.gz SHA256 3dc62de731fd516dfb1032861d9a580d4d0b5b0856beb0f185d06df8e6c26259
2.4.0 (2024-05-07)
librdkafka v2.4.0 is a feature release:
- KIP-848: The Next Generation of the Consumer Rebalance Protocol. Early Access: This should be used only for evaluation and must not be used in production. Features and contract of this KIP might change in future (#4610).
- KIP-467: Augment ProduceResponse error messaging for specific culprit records (#4583).
- KIP-516 Continue partial implementation by adding a metadata cache by topic id and updating the topic id corresponding to the partition name (#4676)
- Upgrade OpenSSL to v3.0.12 (while building from source) with various security fixes, check the release notes.
- Integration tests can be started in KRaft mode and run against any GitHub Kafka branch other than the released versions.
- Fix pipeline inclusion of static binaries (#4666)
- Fix to main loop timeout calculation leading to a tight loop for a max period of 1 ms (#4671).
- Fixed a bug causing duplicate message consumption from a stale fetch start offset in some particular cases (#4636)
- Fix to metadata cache expiration on full metadata refresh (#4677).
- Fix for a wrong error returned on full metadata refresh before joining a consumer group (#4678).
- Fix to metadata refresh interruption (#4679).
- Fix for an undesired partition migration with stale leader epoch (#4680).
- Fix hang in cooperative consumer mode if an assignment is processed while closing the consumer (#4528).
Upgrade considerations
- With KIP 467,
INVALID_MSG
(Java: CorruptRecordExpection) will be retried automatically.INVALID_RECORD
(Java: InvalidRecordException) instead is not retriable and will be set only to the records that caused the error. Rest of records in the batch will fail with the new error code_INVALID_DIFFERENT_RECORD
(Java: KafkaException) and can be retried manually, depending on the application logic (#4583).
Early Access
KIP-848: The Next Generation of the Consumer Rebalance Protocol
- With this new protocol the role of the Group Leader (a member) is removed and the assignment is calculated by the Group Coordinator (a broker) and sent to each member through heartbeats.
The feature is still not production-ready. It's possible to try it in a non-production enviroment.
A guide is available with considerations and steps to follow to test it (#4610).
Fixes
General fixes
- Issues: confluentinc/confluent-kafka-go#981. In librdkafka release pipeline a static build containing libsasl2 could be chosen instead of the alternative one without it. That caused the libsasl2 dependency to be required in confluent-kafka-go v2.1.0-linux-musl-arm64 and v2.3.0-linux-musl-arm64. Solved by correctly excluding the binary configured with that library, when targeting a static build. Happening since v2.0.2, with specified platforms, when using static binaries (#4666).
- Issues: #4684. When the main thread loop was awakened less than 1 ms before the expiration of a timeout, it was serving with a zero timeout, leading to increased CPU usage until the timeout was reached. Happening since 1.x.
- Issues: #4685.
Metadata cache was cleared on full metadata refresh, leading to unnecessary
refreshes and occasional
UNKNOWN_TOPIC_OR_PART
errors. Solved by updating cache for existing or hinted entries instead of clearing them. Happening since 2.1.0 (#4677). - Issues: #4589.
A metadata call before member joins consumer group,
could lead to an
UNKNOWN_TOPIC_OR_PART
error. Solved by updating the consumer group following a metadata refresh only in safe states. Happening since 2.1.0 (#4678). - Issues: #4577. Metadata refreshes without partition leader change could lead to a loop of metadata calls at fixed intervals. Solved by stopping metadata refresh when all existing metadata is non-stale. Happening since 2.3.0 (#4679).
- Issues: #4687. A partition migration could happen, using stale metadata, when the partition was undergoing a validation and being retried because of an error. Solved by doing a partition migration only with a non-stale leader epoch. Happening since 2.1.0 (#4680).
Consumer fixes
- Issues: #4686. In case of subscription change with a consumer using the cooperative assignor it could resume fetching from a previous position. That could also happen if resuming a partition that wasn't paused. Fixed by ensuring that a resume operation is completely a no-op when the partition isn't paused. Happening since 1.x (#4636).
- Issues: #4527.
While using the cooperative assignor, given an assignment is received while closing the consumer
it's possible that it gets stuck in state
WAIT_ASSIGN_CALL
, while the method is converted to a full unassign. Solved by changing state fromWAIT_ASSIGN_CALL
toWAIT_UNASSIGN_CALL
while doing this conversion. Happening since 1.x (#4528).
Checksums
Release asset checksums:
* v2.4.0.zip SHA256 24b30d394fc6ce5535eaa3c356ed9ed9ae4a6c9b4fc9159c322a776786d5dd15
* v2.4.0.tar.gz SHA256 d645e47d961db47f1ead29652606a502bdd2a880c85c1e060e94eea040f1a19a
2.3.0 (2023-10-25)
librdkafka v2.3.0 is a feature release:
- KIP-516
Partial support of topic identifiers. Topic identifiers in metadata response
available through the new
rd_kafka_DescribeTopics
function (#4300, #4451). - KIP-117 Add support for AdminAPI
DescribeCluster()
andDescribeTopics()
(#4240, @jainruchir). - KIP-430: Return authorized operations in Describe Responses. (#4240, @jainruchir).
- KIP-580: Added Exponential Backoff mechanism for
retriable requests with
retry.backoff.ms
as minimum backoff andretry.backoff.max.ms
as the maximum backoff, with 20% jitter (#4422). - KIP-396: completed the implementation with the addition of ListOffsets (#4225).
- Fixed ListConsumerGroupOffsets not fetching offsets for all the topics in a group with Apache Kafka version below 2.4.0.
- Add missing destroy that leads to leaking partition structure memory when there are partition leader changes and a stale leader epoch is received (#4429).
- Fix a segmentation fault when closing a consumer using the cooperative-sticky assignor before the first assignment (#4381).
- Fix for insufficient buffer allocation when allocating rack information (@wolfchimneyrock, #4449).
- Fix for infinite loop of OffsetForLeaderEpoch requests on quick leader changes. (#4433).
- Fix to add leader epoch to control messages, to make sure they're stored for committing even without a subsequent fetch message (#4434).
- Fix for stored offsets not being committed if they lacked the leader epoch (#4442).
- Upgrade OpenSSL to v3.0.11 (while building from source) with various security fixes, check the release notes (#4454, started by @migarc1).
- Fix to ensure permanent errors during offset validation continue being retried and don't cause an offset reset (#4447).
- Fix to ensure max.poll.interval.ms is reset when rd_kafka_poll is called with consume_cb (#4431).
- Fix for idempotent producer fatal errors, triggered after a possibly persisted message state (#4438).
- Fix
rd_kafka_query_watermark_offsets
continuing beyond timeout expiry (#4460). - Fix
rd_kafka_query_watermark_offsets
not refreshing the partition leader after a leader change and subsequentNOT_LEADER_OR_FOLLOWER
error (#4225).
Upgrade considerations
-
retry.backoff.ms
: If it is set greater thanretry.backoff.max.ms
which has the default value of 1000 ms then it is assumes the value ofretry.backoff.max.ms
. To change this behaviour make sure thatretry.backoff.ms
is always less thanretry.backoff.max.ms
. If equal then the backoff will be linear instead of exponential. -
topic.metadata.refresh.fast.interval.ms
: If it is set greater thanretry.backoff.max.ms
which has the default value of 1000 ms then it is assumes the value ofretry.backoff.max.ms
. To change this behaviour make sure thattopic.metadata.refresh.fast.interval.ms
is always less thanretry.backoff.max.ms
. If equal then the backoff will be linear instead of exponential.
Fixes
General fixes
- An assertion failed with insufficient buffer size when allocating rack information on 32bit architectures. Solved by aligning all allocations to the maximum allowed word size (#4449).
- The timeout for
rd_kafka_query_watermark_offsets
was not enforced after making the necessary ListOffsets requests, and thus, it never timed out in case of broker/network issues. Fixed by setting an absolute timeout (#4460).
Idempotent producer fixes
- After a possibly persisted error, such as a disconnection or a timeout, next expected sequence
used to increase, leading to a fatal error if the message wasn't persisted and
the second one in queue failed with an
OUT_OF_ORDER_SEQUENCE_NUMBER
. The error could contain the message "sequence desynchronization" with just one possibly persisted error or "rewound sequence number" in case of multiple errored messages. Solved by treating the possible persisted message as not persisted, and expecting aDUPLICATE_SEQUENCE_NUMBER
error in case it was orNO_ERROR
in case it wasn't, in both cases the message will be considered delivered (#4438).
Consumer fixes
- Stored offsets were excluded from the commit if the leader epoch was less than committed epoch, as it's possible if leader epoch is the default -1. This didn't happen in Python, Go and .NET bindings when stored position was taken from the message. Solved by checking only that the stored offset is greater than committed one, if either stored or committed leader epoch is -1 (#4442).
- If an OffsetForLeaderEpoch request was being retried, and the leader changed while the retry was in-flight, an infinite loop of requests was triggered, because we weren't updating the leader epoch correctly. Fixed by updating the leader epoch before sending the request (#4433).
- During offset validation a permanent error like host resolution failure would cause an offset reset. This isn't what's expected or what the Java implementation does. Solved by retrying even in case of permanent errors (#4447).
- If using
rd_kafka_poll_set_consumer
, along with a consume callback, and then callingrd_kafka_poll
to service the callbacks, would not resetmax.poll.interval.ms.
This was because we were only checkingrk_rep
for consumer messages, while the method to service the queue internally also services the queue forwarded to fromrk_rep
, which isrkcg_q
. Solved by moving themax.poll.interval.ms
check intord_kafka_q_serve
(#4431). - After a leader change a
rd_kafka_query_watermark_offsets
call would continue trying to call ListOffsets on the old leader, if the topic wasn't included in the subscription set, so it started querying the new leader only aftertopic.metadata.refresh.interval.ms
(#4225).
Checksums
Release asset checksums:
* v2.3.0.zip SHA256 15e77455811b3e5d869d6f97ce765b634c7583da188792e2930a2098728e932b
* v2.3.0.tar.gz SHA256 2d49c35c77eeb3d42fa61c43757fcbb6a206daa560247154e60642bcdcc14d12
2.2.0 (2023-07-12)
librdkafka v2.2.0 is a feature release:
- Fix a segmentation fault when subscribing to non-existent topics and using the consume batch functions (#4273).
- Store offset commit metadata in
rd_kafka_offsets_store
(@mathispesch, #4084). - Fix a bug that happens when skipping tags, causing buffer underflow in MetadataResponse (#4278).
- Fix a bug where topic leader is not refreshed in the same metadata call even if the leader is present.
- KIP-881: Add support for rack-aware partition assignment for consumers (#4184, #4291, #4252).
- Fix several bugs with sticky assignor in case of partition ownership changing between members of the consumer group (#4252).
- KIP-368: Allow SASL Connections to Periodically Re-Authenticate (#4301, started by @vctoriawu).
- Avoid treating an OpenSSL error as a permanent error and treat unclean SSL closes as normal ones (#4294).
- Added
fetch.queue.backoff.ms
to the consumer to control how long the consumer backs off next fetch attempt. (@bitemyapp, @edenhill, #2879) - KIP-235: Add DNS alias support for secured connection (#4292).
- KIP-339: IncrementalAlterConfigs API (started by @PrasanthV454, #4110).
- KIP-554: Add Broker-side SCRAM Config API (#4241).
Enhancements
- Added
fetch.queue.backoff.ms
to the consumer to control how long the consumer backs off next fetch attempt. When the pre-fetch queue has exceeded its queuing thresholds:queued.min.messages
andqueued.max.messages.kbytes
it backs off for 1 seconds. If those parameters have to be set too high to hold 1 s of data, this new parameter allows to back off the fetch earlier, reducing memory requirements.
Fixes
General fixes
- Fix a bug that happens when skipping tags, causing buffer underflow in MetadataResponse. This is triggered since RPC version 9 (v2.1.0), when using Confluent Platform, only when racks are set, observers are activated and there is more than one partition. Fixed by skipping the correct amount of bytes when tags are received.
- Avoid treating an OpenSSL error as a permanent error and treat unclean SSL
closes as normal ones. When SSL connections are closed without
close_notify
, in OpenSSL 3.x a new type of error is set and it was interpreted as permanent in librdkafka. It can cause a different issue depending on the RPC. If received when waiting for OffsetForLeaderEpoch response, it triggers an offset reset following the configured policy. Solved by treating SSL errors as transport errors and by setting an OpenSSL flag that allows to treat unclean SSL closes as normal ones. These types of errors can happen it the other side doesn't supportclose_notify
or if there's a TCP connection reset.
Consumer fixes
- In case of multiple owners of a partition with different generations, the sticky assignor would pick the earliest (lowest generation) member as the current owner, which would lead to stickiness violations. Fixed by choosing the latest (highest generation) member.
- In case where the same partition is owned by two members with the same generation, it indicates an issue. The sticky assignor had some code to handle this, but it was non-functional, and did not have parity with the Java assignor. Fixed by invalidating any such partition from the current assignment completely.
Checksums
Release asset checksums:
* v2.2.0.zip SHA256 e9a99476dd326089ce986afd3a5b069ef8b93dbb845bc5157b3d94894de53567
* v2.2.0.tar.gz SHA256 af9a820cbecbc64115629471df7c7cecd40403b6c34bfdbb9223152677a47226
2.1.1 (2023-05-02)
librdkafka v2.1.1 is a maintenance release:
- Avoid duplicate messages when a fetch response is received in the middle of an offset validation request (#4261).
- Fix segmentation fault when subscribing to a non-existent topic and
calling
rd_kafka_message_leader_epoch()
on the polledrkmessage
(#4245). - Fix a segmentation fault when fetching from follower and the partition lease expires while waiting for the result of a list offsets operation (#4254).
- Fix documentation for the admin request timeout, incorrectly stating -1 for infinite timeout. That timeout can't be infinite.
- Fix CMake pkg-config cURL require and use
pkg-config
Requires.private
field (@FantasqueX, @stertingen, #4180). - Fixes certain cases where polling would not keep the consumer in the group or make it rejoin it (#4256).
- Fix to the C++ set_leader_epoch method of TopicPartitionImpl, that wasn't storing the passed value (@pavel-pimenov, #4267).
Fixes
Consumer fixes
- Duplicate messages can be emitted when a fetch response is received in the middle of an offset validation request. Solved by avoiding a restart from last application offset when offset validation succeeds.
- When fetching from follower, if the partition lease expires after 5 minutes,
and a list offsets operation was requested to retrieve the earliest
or latest offset, it resulted in segmentation fault. This was fixed by
allowing threads different from the main one to call
the
rd_kafka_toppar_set_fetch_state
function, given they hold the lock on therktp
. - In v2.1.0, a bug was fixed which caused polling any queue to reset the
max.poll.interval.ms
. Only certain functions were made to reset the timer, but it is possible for the user to obtain the queue with messages from the broker, skipping these functions. This was fixed by encoding information in a queue itself, that, whether polling, resets the timer.
Checksums
Release asset checksums:
* v2.1.1.zip SHA256 3b8a59f71e22a8070e0ae7a6b7ad7e90d39da8fddc41ce6c5d596ee7f5a4be4b
* v2.1.1.tar.gz SHA256 7be1fc37ab10ebdc037d5c5a9b35b48931edafffae054b488faaff99e60e0108
2.1.0 (2023-04-06)
librdkafka v2.1.0 is a feature release:
- KIP-320 Allow fetchers to detect and handle log truncation (#4122).
- Fix a reference count issue blocking the consumer from closing (#4187).
- Fix a protocol issue with ListGroups API, where an extra field was appended for API Versions greater than or equal to 3 (#4207).
- Fix an issue with
max.poll.interval.ms
, where polling any queue would cause the timeout to be reset (#4176). - Fix seek partition timeout, was one thousand times lower than the passed value (#4230).
- Fix multiple inconsistent behaviour in batch APIs during pause or resume operations (#4208). See Consumer fixes section below for more information.
- Update lz4.c from upstream. Fixes CVE-2021-3520 (by @filimonov, #4232).
- Upgrade OpenSSL to v3.0.8 with various security fixes, check the release notes (#4215).
Enhancements
- Added
rd_kafka_topic_partition_get_leader_epoch()
(andset..()
). - Added partition leader epoch APIs:
rd_kafka_topic_partition_get_leader_epoch()
(andset..()
)rd_kafka_message_leader_epoch()
rd_kafka_*assign()
andrd_kafka_seek_partitions()
now supports partitions with a leader epoch set.rd_kafka_offsets_for_times()
will return per-partition leader-epochs.leader_epoch
,stored_leader_epoch
, andcommitted_leader_epoch
added to per-partition statistics.
Fixes
OpenSSL fixes
- Fixed OpenSSL static build not able to use external modules like FIPS provider module.
Consumer fixes
- A reference count issue was blocking the consumer from closing. The problem would happen when a partition is lost, because forcibly unassigned from the consumer or if the corresponding topic is deleted.
- When using
rd_kafka_seek_partitions
, the remaining timeout was converted from microseconds to milliseconds but the expected unit for that parameter is microseconds. - Fixed known issues related to Batch Consume APIs mentioned in v2.0.0 release notes.
- Fixed
rd_kafka_consume_batch()
andrd_kafka_consume_batch_queue()
intermittently updatingapp_offset
andstore_offset
incorrectly when pause or resume was being used for a partition. - Fixed
rd_kafka_consume_batch()
andrd_kafka_consume_batch_queue()
intermittently skipping offsets when pause or resume was being used for a partition.
Known Issues
Consume Batch API
- When
rd_kafka_consume_batch()
andrd_kafka_consume_batch_queue()
APIs are used with any of the seek, pause, resume or rebalancing operation,on_consume
interceptors might be called incorrectly (maybe multiple times) for not consumed messages.
Consume API
- Duplicate messages can be emitted when a fetch response is received in the middle of an offset validation request.
- Segmentation fault when subscribing to a non-existent topic and
calling
rd_kafka_message_leader_epoch()
on the polledrkmessage
.
Checksums
Release asset checksums:
* v2.1.0.zip SHA256 2fe898f9f5e2b287d26c5f929c600e2772403a594a691e0560a2a1f2706edf57
* v2.1.0.tar.gz SHA256 d8e76c4b1cde99e283a19868feaaff5778aa5c6f35790036c5ef44bc5b5187aa
2.0.2 (2023-01-20)
librdkafka v2.0.2 is a bugfix release:
- Fix OpenSSL version in Win32 nuget package (#4152).
Checksums
Release asset checksums:
* v2.0.2.zip SHA256 87010c722111539dc3c258a6be0c03b2d6d4a607168b65992eb0076c647e4e9d
* v2.0.2.tar.gz SHA256 f321bcb1e015a34114c83cf1aa7b99ee260236aab096b85c003170c90a47ca9d
2.0.1 (2023-01-19)
librdkafka v2.0.1 is a bugfix release:
- Fixed nuget package for Linux ARM64 release (#4150).
Checksums
Release asset checksums:
* v2.0.1.zip SHA256 7121df3fad1f72ea1c42dcc4e5367337207a75966216c63e58222c6433c528e0
* v2.0.1.tar.gz SHA256 3670f8d522e77f79f9d09a22387297ab58d1156b22de12ef96e58b7d57fca139
2.0.0 (2023-01-18)
librdkafka v2.0.0 is a feature release:
- KIP-88 OffsetFetch Protocol Update (#3995).
- KIP-222 Add Consumer Group operations to Admin API (started by @lesterfan, #3995).
- KIP-518 Allow listing consumer groups per state (#3995).
- KIP-396 Partially implemented: support for AlterConsumerGroupOffsets (started by @lesterfan, #3995).
- OpenSSL 3.0.x support - the maximum bundled OpenSSL version is now 3.0.7 (previously 1.1.1q).
- Fixes to the transactional and idempotent producer.
Upgrade considerations
OpenSSL 3.0.x
OpenSSL default ciphers
The introduction of OpenSSL 3.0.x in the self-contained librdkafka bundles changes the default set of available ciphers, in particular all obsolete or insecure ciphers and algorithms as listed in the OpenSSL legacy manual page are now disabled by default.
WARNING: These ciphers are disabled for security reasons and it is highly recommended NOT to use them.
Should you need to use any of these old ciphers you'll need to explicitly
enable the legacy
provider by configuring ssl.providers=default,legacy
on the librdkafka client.
OpenSSL engines and providers
OpenSSL 3.0.x deprecates the use of engines, which is being replaced by
providers. As such librdkafka will emit a deprecation warning if
ssl.engine.location
is configured.
OpenSSL providers may be configured with the new ssl.providers
configuration property.
Broker TLS certificate hostname verification
The default value for ssl.endpoint.identification.algorithm
has been
changed from none
(no hostname verification) to https
, which enables
broker hostname verification (to counter man-in-the-middle
impersonation attacks) by default.
To restore the previous behaviour, set ssl.endpoint.identification.algorithm
to none
.
Known Issues
Poor Consumer batch API messaging guarantees
The Consumer Batch APIs rd_kafka_consume_batch()
and rd_kafka_consume_batch_queue()
are not thread safe if rkmessages_size
is greater than 1 and any of the seek,
pause, resume or rebalancing operation is performed in parallel with any of
the above APIs. Some of the messages might be lost, or erroneously returned to the
application, in the above scenario.
It is strongly recommended to use the Consumer Batch APIs and the mentioned operations in sequential order in order to get consistent result.
For rebalancing operation to work in sequencial manner, please set rebalance_cb
configuration property (refer examples/rdkafka_complex_consumer_example.c for the help with the usage) for the consumer.
Enhancements
- Self-contained static libraries can now be built on Linux arm64 (#4005).
- Updated to zlib 1.2.13, zstd 1.5.2, and curl 7.86.0 in self-contained librdkafka bundles.
- Added
on_broker_state_change()
interceptor - The C++ API no longer returns strings by const value, which enables better move optimization in callers.
- Added
rd_kafka_sasl_set_credentials()
API to update SASL credentials. - Setting
allow.auto.create.topics
will no longer give a warning if used by a producer, since that is an expected use case. Improvement in documentation for this property. - Added a
resolve_cb
configuration setting that permits using custom DNS resolution logic. - Added
rd_kafka_mock_broker_error_stack_cnt()
. - The librdkafka.redist NuGet package has been updated to have fewer external dependencies for its bundled librdkafka builds, as everything but cyrus-sasl is now built-in. There are bundled builds with and without linking to cyrus-sasl for maximum compatibility.
- Admin API DescribeGroups() now provides the group instance id for static members KIP-345 (#3995).
Fixes
General fixes
- Windows: couldn't read a PKCS#12 keystore correctly because binary mode wasn't explicitly set and Windows defaults to text mode.
- Fixed memory leak when loading SSL certificates (@Mekk, #3930)
- Load all CA certificates from
ssl.ca.pem
, not just the first one. - Each HTTP request made when using OAUTHBEARER OIDC would leak a small amount of memory.
Transactional producer fixes
- When a PID epoch bump is requested and the producer is waiting to reconnect to the transaction coordinator, a failure in a find coordinator request could cause an assert to fail. This is fixed by retrying when the coordinator is known (#4020).
- Transactional APIs (except
send_offsets_for_transaction()
) that timeout due to low timeout_ms may now be resumed by calling the same API again, as the operation continues in the background. - For fatal idempotent producer errors that may be recovered by bumping the epoch the current transaction must first be aborted prior to the epoch bump. This is now handled correctly, which fixes issues seen with fenced transactional producers on fatal idempotency errors.
- Timeouts for EndTxn requests (transaction commits and aborts) are now automatically retried and the error raised to the application is also a retriable error.
- TxnOffsetCommitRequests were retried immediately upon temporary errors in
send_offsets_to_transactions()
, causing excessive network requests. These retries are now delayed 500ms. - If
init_transactions()
is called with an infinite timeout (-1), the timeout will be limited to 2 *transaction.timeout.ms
. The application may retry and resume the call if a retriable error is returned.
Consumer fixes
- Back-off and retry JoinGroup request if coordinator load is in progress.
- Fix
rd_kafka_consume_batch()
andrd_kafka_consume_batch_queue()
skipping other partitions' offsets intermittently when seek, pause, resume or rebalancing is used for a partition. - Fix
rd_kafka_consume_batch()
andrd_kafka_consume_batch_queue()
intermittently returing incorrect partitions' messages if rebalancing happens during these operations.
Checksums
Release asset checksums:
* v2.0.0.zip SHA256 9d8a8be30ed09daf6c560f402e91db22fcaea11cac18a0d3c0afdbf884df1d4e
* v2.0.0.tar.gz SHA256 f75de3545b3c6cc027306e2df0371aefe1bb8f86d4ec612ed4ebf7bfb2f817cd
1.9.2 (2022-08-01)
librdkafka v1.9.2 is a maintenance release:
- The SASL OAUTHBEAR OIDC POST field was sometimes truncated by one byte (#3192).
- The bundled version of OpenSSL has been upgraded to version 1.1.1q for non-Windows builds. Windows builds remain on OpenSSL 1.1.1n for the time being.
- The bundled version of Curl has been upgraded to version 7.84.0.
Checksums
Release asset checksums:
* v1.9.2.zip SHA256 4ecb0a3103022a7cab308e9fecd88237150901fa29980c99344218a84f497b86
* v1.9.2.tar.gz SHA256 3fba157a9f80a0889c982acdd44608be8a46142270a389008b22d921be1198ad
1.9.1 (2022-07-06)
librdkafka v1.9.1
librdkafka v1.9.1 is a maintenance release:
- The librdkafka.redist NuGet package now contains OSX M1/arm64 builds.
- Self-contained static libraries can now be built on OSX M1 too, thanks to disabling curl's configure runtime check.
Checksums
Release asset checksums:
* v1.9.1.zip SHA256 d3fc2e0bc00c3df2c37c5389c206912842cca3f97dd91a7a97bc0f4fc69f94ce
* v1.9.1.tar.gz SHA256 3a54cf375218977b7af4716ed9738378e37fe400a6c5ddb9d622354ca31fdc79
1.9.0 (2022-06-16)
librdkafka v1.9.0
librdkafka v1.9.0 is a feature release:
- Added KIP-768 OUATHBEARER OIDC support (by @jliunyu, #3560)
- Added KIP-140 Admin API ACL support (by @emasab, #2676)
Upgrade considerations
- Consumer:
rd_kafka_offsets_store()
(et.al) will now return an error for any partition that is not currently assigned (throughrd_kafka_*assign()
). This prevents a race condition where an application would store offsets after the assigned partitions had been revoked (which resets the stored offset), that could cause these old stored offsets to be committed later when the same partitions were assigned to this consumer again - effectively overwriting any committed offsets by any consumers that were assigned the same partitions previously. This would typically result in the offsets rewinding and messages to be reprocessed. As an extra effort to avoid this situation the stored offset is now also reset when partitions are assigned (throughrd_kafka_*assign()
). Applications that explicitly call..offset*_store()
will now need to handle the case whereRD_KAFKA_RESP_ERR__STATE
is returned in the per-partition.err
field - meaning the partition is no longer assigned to this consumer and the offset could not be stored for commit.
Enhancements
- Improved producer queue scheduling. Fixes the performance regression introduced in v1.7.0 for some produce patterns. (#3538, #2912)
- Windows: Added native Win32 IO/Queue scheduling. This removes the internal TCP loopback connections that were previously used for timely queue wakeups.
- Added
socket.connection.setup.timeout.ms
(default 30s). The maximum time allowed for broker connection setups (TCP connection as well as SSL and SASL handshakes) is now limited to this value. This fixes the issue with stalled broker connections in the case of network or load balancer problems. The Java clients has an exponential backoff to this timeout which is limited bysocket.connection.setup.timeout.max.ms
- this was not implemented in librdkafka due to differences in connection handling andERR__ALL_BROKERS_DOWN
error reporting. Having a lower initial connection setup timeout and then increase the timeout for the next attempt would yield possibly false-positiveERR__ALL_BROKERS_DOWN
too early. - SASL OAUTHBEARER refresh callbacks can now be scheduled for execution
on librdkafka's background thread. This solves the problem where an
application has a custom SASL OAUTHBEARER refresh callback and thus needs to
call
rd_kafka_poll()
(et.al.) at least once to trigger the refresh callback before being able to connect to brokers. With the newrd_kafka_conf_enable_sasl_queue()
configuration API andrd_kafka_sasl_background_callbacks_enable()
the refresh callbacks can now be triggered automatically on the librdkafka background thread. rd_kafka_queue_get_background()
now creates the background thread if not already created.- Added
rd_kafka_consumer_close_queue()
andrd_kafka_consumer_closed()
. This allow applications and language bindings to implement asynchronous consumer close. - Bundled zlib upgraded to version 1.2.12.
- Bundled OpenSSL upgraded to 1.1.1n.
- Added
test.mock.broker.rtt
to simulate RTT/latency for mock brokers.
Fixes
General fixes
- Fix various 1 second delays due to internal broker threads blocking on IO
even though there are events to handle.
These delays could be seen randomly in any of the non produce/consume
request APIs, such as
commit_transaction()
,list_groups()
, etc. - Windows: some applications would crash with an error message like
no OPENSSL_Applink()
written to the console ifssl.keystore.location
was configured. This regression was introduced in v1.8.0 due to use of vcpkgs and how keystore file was read. #3554. - Windows 32-bit only: 64-bit atomic reads were in fact not atomic and could
in rare circumstances yield incorrect values.
One manifestation of this issue was the
max.poll.interval.ms
consumer timer expiring even though the application was polling according to profile. Fixed by @WhiteWind (#3815). rd_kafka_clusterid()
would previously fail with timeout if called on cluster with no visible topics (#3620). The clusterid is now returned as soon as metadata has been retrieved.- Fix hang in
rd_kafka_list_groups()
if there are no available brokers to connect to (#3705). - Millisecond timeouts (
timeout_ms
) in various APIs, such asrd_kafka_poll()
, was limited to roughly 36 hours before wrapping. (#3034) - If a metadata request triggered by
rd_kafka_metadata()
or consumer group rebalancing encountered a non-retriable error it would not be propagated to the caller and thus cause a stall or timeout, this has now been fixed. (@aiquestion, #3625) - AdminAPI
DeleteGroups()
andDeleteConsumerGroupOffsets()
: if the given coordinator connection was not up by the time these calls were initiated and the first connection attempt failed then no further connection attempts were performed, ulimately leading to the calls timing out. This is now fixed by keep retrying to connect to the group coordinator until the connection is successful or the call times out. Additionally, the coordinator will be now re-queried once per second until the coordinator comes up or the call times out, to detect change in coordinators. - Mock cluster
rd_kafka_mock_broker_set_down()
would previously accept and then disconnect new connections, it now refuses new connections.
Consumer fixes
rd_kafka_offsets_store()
(et.al) will now return an error for any partition that is not currently assigned (throughrd_kafka_*assign()
). See Upgrade considerations above for more information.rd_kafka_*assign()
will now reset/clear the stored offset. See Upgrade considerations above for more information.seek()
followed bypause()
would overwrite the seeked offset when later callingresume()
. This is now fixed. (#3471). Note: Avoid storing offsets (offsets_store()
) after callingseek()
as this may later interfere with resuming a paused partition, instead store offsets prior to calling seek.- A
ERR_MSG_SIZE_TOO_LARGE
consumer error would previously be raised if the consumer received a maximum sized FetchResponse only containing (transaction) aborted messages with no control messages. The fetching did not stop, but some applications would terminate upon receiving this error. No error is now raised in this case. (#2993) Thanks to @jacobmikesell for providing an application to reproduce the issue. - The consumer no longer backs off the next fetch request (default 500ms) when the parsed fetch response is truncated (which is a valid case). This should speed up the message fetch rate in case of maximum sized fetch responses.
- Fix consumer crash (
assert: rkbuf->rkbuf_rkb
) when parsing malformed JoinGroupResponse consumer group metadata state. - Fix crash (
cant handle op type
) when usingconsume_batch_queue()
(et.al) and an OAUTHBEARER refresh callback was set. The callback is now triggered by the consume call. (#3263) - Fix
partition.assignment.strategy
ordering when multiple strategies are configured. If there is more than one eligible strategy, preference is determined by the configured order of strategies. The partitions are assigned to group members according to the strategy order preference now. (#3818) - Any form of unassign*() (absolute or incremental) is now allowed during consumer close rebalancing and they're all treated as absolute unassigns. (@kevinconaway)
Transactional producer fixes
- Fix message loss in idempotent/transactional producer. A corner case has been identified that may cause idempotent/transactional messages to be lost despite being reported as successfully delivered: During cluster instability a restarting broker may report existing topics as non-existent for some time before it is able to acquire up to date cluster and topic metadata. If an idempotent/transactional producer updates its topic metadata cache from such a broker the producer will consider the topic to be removed from the cluster and thus remove its local partition objects for the given topic. This also removes the internal message sequence number counter for the given partitions. If the producer later receives proper topic metadata for the cluster the previously "removed" topics will be rediscovered and new partition objects will be created in the producer. These new partition objects, with no knowledge of previous incarnations, would start counting partition messages at zero again. If new messages were produced for these partitions by the same producer instance, the same message sequence numbers would be sent to the broker. If the broker still maintains state for the producer's PID and Epoch it could deem that these messages with reused sequence numbers had already been written to the log and treat them as legit duplicates. This would seem to the producer that these new messages were successfully written to the partition log by the broker when they were in fact discarded as duplicates, leading to silent message loss. The fix included in this release is to save the per-partition idempotency state when a partition is removed, and then recover and use that saved state if the partition comes back at a later time.
- The transactional producer would retry (re)initializing its PID if a
PRODUCER_FENCED
error was returned from the broker (added in Apache Kafka 2.8), which could cause the producer to seemingly hang. This error code is now correctly handled by raising a fatal error. - If the given group coordinator connection was not up by the time
send_offsets_to_transactions()
was called, and the first connection attempt failed then no further connection attempts were performed, ulimately leading tosend_offsets_to_transactions()
timing out, and possibly also the transaction timing out on the transaction coordinator. This is now fixed by keep retrying to connect to the group coordinator until the connection is successful or the call times out. Additionally, the coordinator will be now re-queried once per second until the coordinator comes up or the call times out, to detect change in coordinators.
Producer fixes
- Improved producer queue wakeup scheduling. This should significantly decrease the number of wakeups and thus syscalls for high message rate producers. (#3538, #2912)
- The logic for enforcing that
message.timeout.ms
is greather than an explicitly configuredlinger.ms
was incorrect and instead of erroring out early the lingering time was automatically adjusted to the message timeout, ignoring the configuredlinger.ms
. This has now been fixed so that an error is returned when instantiating the producer. Thanks to @larry-cdn77 for analysis and test-cases. (#3709)
Checksums
Release asset checksums:
* v1.9.0.zip SHA256 a2d124cfb2937ec5efc8f85123dbcfeba177fb778762da506bfc5a9665ed9e57
* v1.9.0.tar.gz SHA256 59b6088b69ca6cf278c3f9de5cd6b7f3fd604212cd1c59870bc531c54147e889
1.6.2 (2021-11-25)
librdkafka v1.6.2
librdkafka v1.6.2 is a maintenance release with the following backported fixes:
- Upon quick repeated leader changes the transactional producer could receive
an
OUT_OF_ORDER_SEQUENCE
error from the broker, which triggered an Epoch bump on the producer resulting in an InitProducerIdRequest being sent to the transaction coordinator in the middle of a transaction. This request would start a new transaction on the coordinator, but the producer would still think (erroneously) it was in the current transaction. Any messages produced in the current transaction prior to this event would be silently lost when the application committed the transaction, leading to message loss. To avoid message loss a fatal error is now raised. This fix is specific to v1.6.x. librdkafka v1.8.x implements a recoverable error state instead. #3575. - The transactional producer could stall during a transaction if the transaction coordinator changed while adding offsets to the transaction (send_offsets_to_transaction()). This stall lasted until the coordinator connection went down, the transaction timed out, transaction was aborted, or messages were produced to a new partition, whichever came first. #3571.
- librdkafka's internal timers would not start if the timeout was set to 0, which would result in some timeout operations not being enforced correctly, e.g., the transactional producer API timeouts. These timers are now started with a timeout of 1 microsecond.
- Force address resolution if the broker epoch changes (#3238).
Checksums
Release asset checksums:
* v1.6.2.zip SHA256 1d389a98bda374483a7b08ff5ff39708f5a923e5add88b80b71b078cb2d0c92e
* v1.6.2.tar.gz SHA256 b9be26c632265a7db2fdd5ab439f2583d14be08ab44dc2e33138323af60c39db
1.8.2 (2021-10-18)
librdkafka v1.8.2
librdkafka v1.8.2 is a maintenance release.
Enhancements
- Added
ssl.ca.pem
to add CA certificate by PEM string. (#2380) - Prebuilt binaries for Mac OSX now contain statically linked OpenSSL v1.1.1l. Previously the OpenSSL version was either v1.1.1 or v1.0.2 depending on build type.
Fixes
- The
librdkafka.redist
1.8.0 package had two flaws: - the linux-arm64 .so build was a linux-x64 build.
- the included Windows MSVC 140 runtimes for x64 were infact x86. The release script has been updated to verify the architectures of provided artifacts to avoid this happening in the future.
- Prebuilt binaries for Mac OSX Sierra (10.12) and older are no longer provided. This affects confluent-kafka-go.
- Some of the prebuilt binaries for Linux were built on Ubuntu 14.04, these builds are now performed on Ubuntu 16.04 instead. This may affect users on ancient Linux distributions.
- It was not possible to configure
ssl.ca.location
on OSX, the property would automatically revert back toprobe
(default value). This regression was introduced in v1.8.0. (#3566) - librdkafka's internal timers would not start if the timeout was set to 0, which would result in some timeout operations not being enforced correctly, e.g., the transactional producer API timeouts. These timers are now started with a timeout of 1 microsecond.
Transactional producer fixes
- Upon quick repeated leader changes the transactional producer could receive
an
OUT_OF_ORDER_SEQUENCE
error from the broker, which triggered an Epoch bump on the producer resulting in an InitProducerIdRequest being sent to the transaction coordinator in the middle of a transaction. This request would start a new transaction on the coordinator, but the producer would still think (erroneously) it was in current transaction. Any messages produced in the current transaction prior to this event would be silently lost when the application committed the transaction, leading to message loss. This has been fixed by setting the Abortable transaction error state in the producer. #3575. - The transactional producer could stall during a transaction if the transaction coordinator changed while adding offsets to the transaction (send_offsets_to_transaction()). This stall lasted until the coordinator connection went down, the transaction timed out, transaction was aborted, or messages were produced to a new partition, whichever came first. #3571.
Checksums
Release asset checksums:
* v1.8.2.zip SHA256 8b03d8b650f102f3a6a6cff6eedc29b9e2f68df9ba7e3c0f3fb00838cce794b8
* v1.8.2.tar.gz SHA256 6a747d293a7a4613bd2897e28e8791476fbe1ae7361f2530a876e0fd483482a6
Note: there was no v1.8.1 librdkafka release
1.8.0 (2021-09-16)
librdkafka v1.8.0
librdkafka v1.8.0 is a security release:
- Upgrade bundled zlib version from 1.2.8 to 1.2.11 in the
librdkafka.redist
NuGet package. The updated zlib version fixes CVEs: CVE-2016-9840, CVE-2016-9841, CVE-2016-9842, CVE-2016-9843 See https://github.com/edenhill/librdkafka/issues/2934 for more information. - librdkafka now uses vcpkg for up-to-date Windows
dependencies in the
librdkafka.redist
NuGet package: OpenSSL 1.1.1l, zlib 1.2.11, zstd 1.5.0. - The upstream dependency (OpenSSL, zstd, zlib) source archive checksums are
now verified when building with
./configure --install-deps
. These builds are used by the librdkafka builds bundled with confluent-kafka-go, confluent-kafka-python and confluent-kafka-dotnet.
Enhancements
- Producer
flush()
now overrides thelinger.ms
setting for the duration of theflush()
call, effectively triggering immediate transmission of queued messages. (#3489)
Fixes
General fixes
- Correctly detect presence of zlib via compilation check. (Chris Novakovic)
ERR__ALL_BROKERS_DOWN
is no longer emitted when the coordinator connection goes down, only when all standard named brokers have been tried. This fixes the issue withERR__ALL_BROKERS_DOWN
being triggered onconsumer_close()
. It is also now only emitted if the connection was fully up (past handshake), and not just connected.rd_kafka_query_watermark_offsets()
,rd_kafka_offsets_for_times()
,consumer_lag
metric, andauto.offset.reset
now honourisolation.level
and will return the Last Stable Offset (LSO) whenisolation.level
is set toread_committed
(default), rather than the uncommitted high-watermark when it is set toread_uncommitted
. (#3423)- SASL GSSAPI is now usable when
sasl.kerberos.min.time.before.relogin
is set to 0 - which disables ticket refreshes (by @mpekalski, #3431). - Rename internal crc32c() symbol to rd_crc32c() to avoid conflict with other static libraries (#3421).
txidle
andrxidle
in the statistics object was emitted as 18446744073709551615 when no idle was known. -1 is now emitted instead. (#3519)
Consumer fixes
- Automatically retry offset commits on
ERR_REQUEST_TIMED_OUT
,ERR_COORDINATOR_NOT_AVAILABLE
, andERR_NOT_COORDINATOR
(#3398). Offset commits will be retried twice. - Timed auto commits did not work when only using assign() and not subscribe(). This regression was introduced in v1.7.0.
- If the topics matching the current subscription changed (or the application updated the subscription) while there was an outstanding JoinGroup or SyncGroup request, an additional request would sometimes be sent before handling the response of the first. This in turn lead to internal state issues that could cause a crash or malbehaviour. The consumer will now wait for any outstanding JoinGroup or SyncGroup responses before re-joining the group.
auto.offset.reset
could previously be triggered by temporary errors, such as disconnects and timeouts (after the two retries are exhausted). This is now fixed so that the auto offset reset policy is only triggered for permanent errors.- The error that triggers
auto.offset.reset
is now logged to help the application owner identify the reason of the reset. - If a rebalance takes longer than a consumer's
session.timeout.ms
, the consumer will remain in the group as long as it receives heartbeat responses from the broker.
Admin fixes
DeleteRecords()
could crash if one of the underlying requests (for a given partition leader) failed at the transport level (e.g., timeout). (#3476).
Checksums
Release asset checksums:
* v1.8.0.zip SHA256 4b173f759ea5fdbc849fdad00d3a836b973f76cbd3aa8333290f0398fd07a1c4
* v1.8.0.tar.gz SHA256 93b12f554fa1c8393ce49ab52812a5f63e264d9af6a50fd6e6c318c481838b7f
1.7.0 (2021-05-10)
librdkafka v1.7.0
librdkafka v1.7.0 is feature release:
- KIP-360 - Improve reliability of transactional producer. Requires Apache Kafka 2.5 or later.
- OpenSSL Engine support (
ssl.engine.location
) by @adinigam and @ajbarb.
Enhancements
- Added
connections.max.idle.ms
to automatically close idle broker connections. This feature is disabled by default unlessbootstrap.servers
contains the stringazure
in which case the default is set to <4 minutes to improve connection reliability and circumvent limitations with the Azure load balancers (see #3109 for more information). - Bumped to OpenSSL 1.1.1k in binary librdkafka artifacts.
- The binary librdkafka artifacts for Alpine are now using Alpine 3.12.
- Improved static librdkafka Windows builds using MinGW (@neptoess, #3130).
Upgrade considerations
- The C++
oauthbearer_token_refresh_cb()
was missing aHandle *
argument that has now been added. This is a breaking change but the original function signature is considered a bug. This change only affects C++ OAuth developers. - KIP-735 The consumer
session.timeout.ms
default was changed from 10 to 45 seconds to make consumer groups more robust and less sensitive to temporary network and cluster issues. - Statistics:
consumer_lag
is now using thecommitted_offset
, while the newconsumer_lag_stored
is usingstored_offset
(offset to be committed). This is more correct than the previousconsumer_lag
which was using eithercommitted_offset
orapp_offset
(last message passed to application).
Fixes
General fixes
- Fix accesses to freed metadata cache mutexes on client termination (#3279)
- There was a race condition on receiving updated metadata where a broker id
update (such as bootstrap to proper broker transformation) could finish after
the topic metadata cache was updated, leading to existing brokers seemingly
being not available.
One occurrence of this issue was query_watermark_offsets() that could return
ERR__UNKNOWN_PARTITION
for existing partitions shortly after the client instance was created. - The OpenSSL context is now initialized with
TLS_client_method()
(on OpenSSL >= 1.1.0) instead of the deprecated and outdatedSSLv23_client_method()
. - The initial cluster connection on client instance creation could sometimes
be delayed up to 1 second if a
group.id
ortransactional.id
was configured (#3305). - Speed up triggering of new broker connections in certain cases by exiting the broker thread io/op poll loop when a wakeup op is received.
- SASL GSSAPI: The Kerberos kinit refresh command was triggered from
rd_kafka_new()
which made this call blocking if the refresh command was taking long. The refresh is now performed by the background rdkafka main thread. - Fix busy-loop (100% CPU on the broker threads) during the handshake phase of an SSL connection.
- Disconnects during SSL handshake are now propagated as transport errors rather than SSL errors, since these disconnects are at the transport level (e.g., incorrect listener, flaky load balancer, etc) and not due to SSL issues.
- Increment metadata fast refresh interval backoff exponentially (@ajbarb, #3237).
- Unthrottled requests are no longer counted in the
brokers[].throttle
statistics object. - Log CONFWARN warning when global topic configuration properties
are overwritten by explicitly setting a
default_topic_conf
.
Consumer fixes
- If a rebalance happened during a
consume_batch..()
call the already accumulated messages for revoked partitions were not purged, which would pass messages to the application for partitions that were no longer owned by the consumer. Fixed by @jliunyu. #3340. - Fix balancing and reassignment issues with the cooperative-sticky assignor. #3306.
- Fix incorrect detection of first rebalance in sticky assignor (@hallfox).
- Aborted transactions with no messages produced to a partition could
cause further successfully committed messages in the same Fetch response to
be ignored, resulting in consumer-side message loss.
A log message along the lines
Abort txn ctrl msg bad order at offset 7501: expected before or at 7702: messages in aborted transactions may be delivered to the application
would be seen. This is a rare occurrence where a transactional producer would register with the partition but not produce any messages before aborting the transaction. - The consumer group deemed cached metadata up to date by checking
topic.metadata.refresh.interval.ms
: if this property was set too low it would cause cached metadata to be unusable and new metadata to be fetched, which could delay the time it took for a rebalance to settle. It now correctly usesmetadata.max.age.ms
instead. - The consumer group timed auto commit would attempt commits during rebalances, which could result in "Illegal generation" errors. This is now fixed, the timed auto committer is only employed in the steady state when no rebalances are taking places. Offsets are still auto committed when partitions are revoked.
- Retriable FindCoordinatorRequest errors are no longer propagated to the application as they are retried automatically.
- Fix rare crash (assert
rktp_started
) on consumer termination (introduced in v1.6.0). - Fix unaligned access and possibly corrupted snappy decompression when building with MSVC (@azat)
- A consumer configured with the
cooperative-sticky
assignor did not actively Leave the group on unsubscribe(). This delayed the rebalance for the remaining group members by up tosession.timeout.ms
. - The current subscription list was sometimes leaked when unsubscribing.
Producer fixes
- The timeout value of
flush()
was not respected when delivery reports were scheduled as events (such as for confluent-kafka-go) rather than callbacks. - There was a race conditition in
purge()
which could cause newly created partition objects, or partitions that were changing leaders, to not have their message queues purged. This could causeabort_transaction()
to time out. This issue is now fixed. - In certain high-thruput produce rate patterns producing could stall for
1 second, regardless of
linger.ms
, due to rate-limiting of internal queue wakeups. This is now fixed by not rate-limiting queue wakeups but instead limiting them to one wakeup per queue reader poll. #2912.
Transactional Producer fixes
- KIP-360: Fatal Idempotent producer errors are now recoverable by the
transactional producer and will raise a
txn_requires_abort()
error. - If the cluster went down between
produce()
andcommit_transaction()
and before any partitions had been registered with the coordinator, the messages would time out but the commit would succeed because nothing had been sent to the coordinator. This is now fixed. - If the current transaction failed while
commit_transaction()
was checking the current transaction state an invalid state transaction could occur which in turn would trigger a assertion crash. This issue showed up as "Invalid txn state transition: .." crashes, and is now fixed by properly synchronizing both checking and transition of state.
1.6.1 (2021-02-24)
librdkafka v1.6.1
librdkafka v1.6.1 is a maintenance release.
Upgrade considerations
- Fatal idempotent producer errors are now also fatal to the transactional producer. This is a necessary step to maintain data integrity prior to librdkafka supporting KIP-360. Applications should check any transactional API errors for the is_fatal flag and decommission the transactional producer if the flag is set.
- The consumer error raised by
auto.offset.reset=error
now has error-code set toERR__AUTO_OFFSET_RESET
to allow an application to differentiate between auto offset resets and other consumer errors.
Fixes
General fixes
- Admin API and transactional
send_offsets_to_transaction()
coordinator requests, such as TxnOffsetCommitRequest, could in rare cases be sent multiple times which could cause a crash. ssl.ca.location=probe
is now enabled by default on Mac OSX since the librdkafka-bundled OpenSSL might not have the same default CA search paths as the system or brew installed OpenSSL. Probing scans all known locations.
Transactional Producer fixes
- Fatal idempotent producer errors are now also fatal to the transactional producer.
- The transactional producer could crash if the transaction failed while
send_offsets_to_transaction()
was called. - Group coordinator requests for transactional
send_offsets_to_transaction()
calls would leak memory if the underlying request was attempted to be sent after the transaction had failed. - When gradually producing to multiple partitions (resulting in multiple underlying AddPartitionsToTxnRequests) sub-sequent partitions could get stuck in pending state under certain conditions. These pending partitions would not send queued messages to the broker and eventually trigger message timeouts, failing the current transaction. This is now fixed.
- Committing an empty transaction (no messages were produced and no offsets were sent) would previously raise a fatal error due to invalid state on the transaction coordinator. We now allow empty/no-op transactions to be committed.
Consumer fixes
- The consumer will now retry indefinitely (or until the assignment is changed)
to retrieve committed offsets. This fixes the issue where only two retries
were attempted when outstanding transactions were blocking OffsetFetch
requests with
ERR_UNSTABLE_OFFSET_COMMIT
. #3265
1.6.0 (2021-01-26)
librdkafka v1.6.0
librdkafka v1.6.0 is feature release:
- KIP-429 Incremental rebalancing with sticky consumer group partition assignor (KIP-54) (by @mhowlett).
- KIP-480 Sticky producer partitioning (
sticky.partitioning.linger.ms
) - achieves higher throughput and lower latency through sticky selection of random partition (by @abbycriswell). - AdminAPI: Add support for
DeleteRecords()
,DeleteGroups()
andDeleteConsumerGroupOffsets()
(by @gridaphobe) - KIP-447 Producer scalability for exactly once semantics - allows a single transactional producer to be used for multiple input partitions. Requires Apache Kafka 2.5 or later.
- Transactional producer fixes and improvements, see Transactional Producer fixes below.
- The librdkafka.redist NuGet package now supports Linux ARM64/Aarch64.
Upgrade considerations
- Sticky producer partitioning (
sticky.partitioning.linger.ms
) is enabled by default (10 milliseconds) which affects the distribution of randomly partitioned messages, where previously these messages would be evenly distributed over the available partitions they are now partitioned to a single partition for the duration of the sticky time (10 milliseconds by default) before a new random sticky partition is selected. - The new KIP-447 transactional producer scalability guarantees are only supported on Apache Kafka 2.5 or later, on earlier releases you will need to use one producer per input partition for EOS. This limitation is not enforced by the producer or broker.
- Error handling for the transactional producer has been improved, see the Transactional Producer fixes below for more information.
Known issues
- The Transactional Producer's API timeout handling is inconsistent with the
underlying protocol requests, it is therefore strongly recommended that
applications call
rd_kafka_commit_transaction()
andrd_kafka_abort_transaction()
with thetimeout_ms
parameter set to-1
, which will use the remaining transaction timeout.
Enhancements
- KIP-107, KIP-204: AdminAPI: Added
DeleteRecords()
(by @gridaphobe). - KIP-229: AdminAPI: Added
DeleteGroups()
(by @gridaphobe). - KIP-496: AdminAPI: Added
DeleteConsumerGroupOffsets()
. - KIP-464: AdminAPI: Added support for broker-side default partition count
and replication factor for
CreateTopics()
. - Windows: Added
ssl.ca.certificate.stores
to specify a list of Windows Certificate Stores to read CA certificates from, e.g.,CA,Root
.Root
remains the default store. - Use reentrant
rand_r()
on supporting platforms which decreases lock contention (@azat). - Added
assignor
debug context for troubleshooting consumer partition assignments. - Updated to OpenSSL v1.1.1i when building dependencies.
- Update bundled lz4 (used when
./configure --disable-lz4-ext
) to v1.9.3 which has vast performance improvements. - Added
rd_kafka_conf_get_default_topic_conf()
to retrieve the default topic configuration object from a global configuration object. - Added
conf
debugging context todebug
- shows set configuration properties on client and topic instantiation. Sensitive properties are redacted. - Added
rd_kafka_queue_yield()
to cancel a blocking queue call. - Will now log a warning when multiple ClusterIds are seen, which is an indication that the client might be erroneously configured to connect to multiple clusters which is not supported.
- Added
rd_kafka_seek_partitions()
to seek multiple partitions to per-partition specific offsets.
Fixes
General fixes
- Fix a use-after-free crash when certain coordinator requests were retried.
- The C++
oauthbearer_set_token()
function would callfree()
on anew
-created pointer, possibly leading to crashes or heap corruption (#3194)
Consumer fixes
- The consumer assignment and consumer group implementations have been decoupled, simplified and made more strict and robust. This will sort out a number of edge cases for the consumer where the behaviour was previously undefined.
- Partition fetch state was not set to STOPPED if OffsetCommit failed.
- The session timeout is now enforced locally also when the coordinator connection is down, which was not previously the case.
Transactional Producer fixes
- Transaction commit or abort failures on the broker, such as when the producer was fenced by a newer instance, were not propagated to the application resulting in failed commits seeming successful. This was a critical race condition for applications that had a delay after producing messages (or sendings offsets) before committing or aborting the transaction. This issue has now been fixed and test coverage improved.
- The transactional producer API would return
RD_KAFKA_RESP_ERR__STATE
when API calls were attempted after the transaction had failed, we now try to return the error that caused the transaction to fail in the first place, such asRD_KAFKA_RESP_ERR__FENCED
when the producer has been fenced, orRD_KAFKA_RESP_ERR__TIMED_OUT
when the transaction has timed out. - Transactional producer retry count for transactional control protocol
requests has been increased from 3 to infinite, retriable errors
are now automatically retried by the producer until success or the
transaction timeout is exceeded. This fixes the case where
rd_kafka_send_offsets_to_transaction()
would fail the current transaction into an abortable state whenCONCURRENT_TRANSACTIONS
was returned by the broker (which is a transient error) and the 3 retries were exhausted.
Producer fixes
- Calling
rd_kafka_topic_new()
with a topic config object withmessage.timeout.ms
set could sometimes adjust the globallinger.ms
property (if not explicitly configured) which was not desired, this is now fixed and the auto adjustment is only done based on thedefault_topic_conf
at producer creation. rd_kafka_flush()
could previously returnRD_KAFKA_RESP_ERR__TIMED_OUT
just as the timeout was reached if the messages had been flushed but there were now no more messages. This has been fixed.
Checksums
Release asset checksums:
* v1.6.0.zip SHA256 af6f301a1c35abb8ad2bb0bab0e8919957be26c03a9a10f833c8f97d6c405aa8
* v1.6.0.tar.gz SHA256 3130cbd391ef683dc9acf9f83fe82ff93b8730a1a34d0518e93c250929be9f6b
1.5.3 (2020-12-09)
librdkafka v1.5.3
librdkafka v1.5.3 is a maintenance release.
Upgrade considerations
- CentOS 6 is now EOL and is no longer included in binary librdkafka packages, such as NuGet.
Fixes
General fixes
- Fix a use-after-free crash when certain coordinator requests were retried.
Consumer fixes
- Consumer would not filter out messages for aborted transactions if the messages were compressed (#3020).
- Consumer destroy without prior
close()
could hang in certain cgrp states (@gridaphobe, #3127). - Fix possible null dereference in
Message::errstr()
(#3140). - The
roundrobin
partition assignment strategy could get stuck in an endless loop or generate uneven assignments in case the group members had asymmetric subscriptions (e.g., c1 subscribes to t1,t2 while c2 subscribes to t2,t3). (#3159)
Checksums
Release asset checksums:
* v1.5.3.zip SHA256 3f24271232a42f2d5ac8aab3ab1a5ddbf305f9a1ae223c840d17c221d12fe4c1
* v1.5.3.tar.gz SHA256 2105ca01fef5beca10c9f010bc50342b15d5ce6b73b2489b012e6d09a008b7bf
1.5.2 (2020-10-20)
librdkafka v1.5.2
librdkafka v1.5.2 is a maintenance release.
Upgrade considerations
- The default value for the producer configuration property
retries
has been increased from 2 to infinity, effectively limiting Produce retries to onlymessage.timeout.ms
. As the reasons for the automatic internal retries vary (various broker error codes as well as transport layer issues), it doesn't make much sense to limit the number of retries for retriable errors, but instead only limit the retries based on the allowed time to produce a message. - The default value for the producer configuration property
request.timeout.ms
has been increased from 5 to 30 seconds to match the Apache Kafka Java producer default. This change yields increased robustness for broker-side congestion.
Enhancements
- The generated
CONFIGURATION.md
(throughrd_kafka_conf_properties_show())
) now include all properties and values, regardless if they were included in the build, and setting a disabled property or value throughrd_kafka_conf_set()
now returnsRD_KAFKA_CONF_INVALID
and provides a more useful error string saying why the property can't be set. - Consumer configs on producers and vice versa will now be logged with warning messages on client instantiation.
Fixes
Security fixes
- There was an incorrect call to zlib's
inflateGetHeader()
with unitialized memory pointers that could lead to the GZIP header of a fetched message batch to be copied to arbitrary memory. This function call has now been completely removed since the result was not used. Reported by Ilja van Sprundel.
General fixes
rd_kafka_topic_opaque()
(used by the C++ API) would cause object refcounting issues when used on light-weight (error-only) topic objects such as consumer errors (#2693).- Handle name resolution failures when formatting IP addresses in error logs, and increase printed hostname limit to ~256 bytes (was ~60).
- Broker sockets would be closed twice (thus leading to potential race
condition with fd-reuse in other threads) if a custom
socket_cb
would return error.
Consumer fixes
- The
roundrobin
partition.assignment.strategy
could crash (assert) for certain combinations of members and partitions. This is a regression in v1.5.0. (#3024) - The C++
KafkaConsumer
destructor did not destroy the underlying Crd_kafka_t
instance, causing a leak ifclose()
was not used. - Expose rich error strings for C++ Consumer
Message->errstr()
. - The consumer could get stuck if an outstanding commit failed during rebalancing (#2933).
- Topic authorization errors during fetching are now reported only once (#3072).
Producer fixes
- Topic authorization errors are now properly propagated for produced messages,
both through delivery reports and as
ERR_TOPIC_AUTHORIZATION_FAILED
return value fromproduce*()
(#2215) - Treat cluster authentication failures as fatal in the transactional producer (#2994).
- The transactional producer code did not properly reference-count partition objects which could in very rare circumstances lead to a use-after-free bug if a topic was deleted from the cluster when a transaction was using it.
ERR_KAFKA_STORAGE_ERROR
is now correctly treated as a retriable produce error (#3026).- Messages that timed out locally would not fail the ongoing transaction. If the application did not take action on failed messages in its delivery report callback and went on to commit the transaction, the transaction would be successfully committed, simply omitting the failed messages.
- EndTxnRequests (sent on commit/abort) are only retried in allowed states (#3041). Previously the transaction could hang on commit_transaction() if an abortable error was hit and the EndTxnRequest was to be retried.
Note: there was no v1.5.1 librdkafka release
Checksums
Release asset checksums:
* v1.5.2.zip SHA256 de70ebdb74c7ef8c913e9a555e6985bcd4b96eb0c8904572f3c578808e0992e1
* v1.5.2.tar.gz SHA256 ca3db90d04ef81ca791e55e9eed67e004b547b7adedf11df6c24ac377d4840c6
1.5.0 (2020-07-20)
librdkafka v1.5.0
The v1.5.0 release brings usability improvements, enhancements and fixes to librdkafka.
Enhancements
- Improved broker connection error reporting with more useful information and hints on the cause of the problem.
- Consumer: Propagate errors when subscribing to unavailable topics (#1540)
- Producer: Add
batch.size
producer configuration property (#638) - Add
topic.metadata.propagation.max.ms
to allow newly manually created topics to be propagated throughout the cluster before reporting them as non-existent. This fixes race issues where CreateTopics() is quickly followed by produce(). - Prefer least idle connection for periodic metadata refreshes, et.al., to allow truly idle connections to time out and to avoid load-balancer-killed idle connection errors (#2845)
- Added
rd_kafka_event_debug_contexts()
to get the debug contexts for a debug log line (by @wolfchimneyrock). - Added Test scenarios which define the cluster configuration.
- Added MinGW-w64 builds (@ed-alertedh, #2553)
./configure --enable-XYZ
now requires the XYZ check to pass, and--disable-XYZ
disables the feature altogether (@benesch)- Added
rd_kafka_produceva()
which takes an array of produce arguments for situations where the existingrd_kafka_producev()
va-arg approach can't be used. - Added
rd_kafka_message_broker_id()
to see the broker that a message was produced or fetched from, or an error was associated with. - Added RTT/delay simulation to mock brokers.
Upgrade considerations
- Subscribing to non-existent and unauthorized topics will now propagate
errors
RD_KAFKA_RESP_ERR_UNKNOWN_TOPIC_OR_PART
andRD_KAFKA_RESP_ERR_TOPIC_AUTHORIZATION_FAILED
to the application through the standard consumer error (the err field in the message object). - Consumer will no longer trigger auto creation of topics,
allow.auto.create.topics=true
may be used to re-enable the old deprecated functionality. - The default consumer pre-fetch queue threshold
queued.max.messages.kbytes
has been decreased from 1GB to 64MB to avoid excessive network usage for low and medium throughput consumer applications. High throughput consumer applications may need to manually set this property to a higher value. - The default consumer Fetch wait time has been increased from 100ms to 500ms to avoid excessive network usage for low throughput topics.
- If OpenSSL is linked statically, or
ssl.ca.location=probe
is configured, librdkafka will probe known CA certificate paths and automatically use the first one found. This should alleviate the need to configuressl.ca.location
when the statically linked OpenSSL's OPENSSLDIR differs from the system's CA certificate path. - The heuristics for handling Apache Kafka < 0.10 brokers has been removed to
improve connection error handling for modern Kafka versions.
Users on Brokers 0.9.x or older should already be configuring
api.version.request=false
andbroker.version.fallback=...
so there should be no functional change. - The default producer batch accumulation time,
linger.ms
, has been changed from 0.5ms to 5ms to improve batch sizes and throughput while reducing the per-message protocol overhead. Applications that require lower produce latency than 5ms will need to manually setlinger.ms
to a lower value. - librdkafka's build tooling now requires Python 3.x (python3 interpreter).
Fixes
General fixes
- The client could crash in rare circumstances on ApiVersion or SaslHandshake request timeouts (#2326)
./configure --LDFLAGS='a=b, c=d'
with arguments containing = are now supported (by @sky92zwq)../configure
arguments now take precedence over cachedconfigure
variables from previous invocation.- Fix theoretical crash on coord request failure.
- Unknown partition error could be triggered for existing partitions when additional partitions were added to a topic (@benesch, #2915)
- Quickly refresh topic metadata for desired but non-existent partitions. This will speed up the initial discovery delay when new partitions are added to an existing topic (#2917).
Consumer fixes
- The roundrobin partition assignor could crash if subscriptions where asymmetrical (different sets from different members of the group). Thanks to @ankon and @wilmai for identifying the root cause (#2121).
- The consumer assignors could ignore some topics if there were more subscribed topics than consumers in taking part in the assignment.
- The consumer would connect to all partition leaders of a topic even for partitions that were not being consumed (#2826).
- Initial consumer group joins should now be a couple of seconds quicker thanks expedited query intervals (@benesch).
- Fix crash and/or inconsistent subscriptions when using multiple consumers (in the same process) with wildcard topics on Windows.
- Don't propagate temporary offset lookup errors to application.
- Immediately refresh topic metadata when partitions are reassigned to other
brokers, avoiding a fetch stall of up to
topic.metadata.refresh.interval.ms
. (#2955) - Memory for batches containing control messages would not be freed when using the batch consume APIs (@pf-qiu, #2990).
Producer fixes
- Proper locking for transaction state in EndTxn handler.
Checksums
Release asset checksums:
* v1.5.0.zip SHA256 76a1e83d643405dd1c0e3e62c7872b74e3a96c52be910233e8ec02d501fa33c8
* v1.5.0.tar.gz SHA256 f7fee59fdbf1286ec23ef0b35b2dfb41031c8727c90ced6435b8cf576f23a656
1.4.4 (2020-06-20)
librdkafka v1.4.4
v1.4.4 is a maintenance release with the following fixes and enhancements:
- Transactional producer could crash on request timeout due to dereferencing NULL pointer of non-existent response object.
- Mark
rd_kafka_send_offsets_to_transaction()
CONCURRENT_TRANSACTION (et.al) errors as retriable. - Fix crash on transactional coordinator FindCoordinator request failure.
- Minimize broker re-connect delay when broker's connection is needed to send requests.
socket.timeout.ms
was ignored whentransactional.id
was set.- Added RTT/delay simulation to mock brokers.
Note: there was no v1.4.3 librdkafka release
1.4.2 (2020-05-06)
librdkafka v1.4.2
v1.4.2 is a maintenance release with the following fixes and enhancements:
- Fix produce/consume hang after partition goes away and comes back, such as when a topic is deleted and re-created (regression in v1.3.0).
- Consumer: Reset the stored offset when partitions are un-assign()ed (fixes #2782). This fixes the case where a manual offset-less commit() or the auto-committer would commit a stored offset from a previous assignment before a new message was consumed by the application.
- Probe known CA cert paths and set default
ssl.ca.location
accordingly if OpenSSL is statically linked orssl.ca.location
is set toprobe
. - Per-partition OffsetCommit errors were unhandled (fixes #2791)
- Seed the PRNG (random number generator) by default, allow application to override with
enable.random.seed=false
(#2795) - Fix stack overwrite (of 1 byte) when SaslHandshake MechCnt is zero
- Align bundled c11 threads (tinycthreads) constants to glibc and musl (#2681)
- Fix return value of rd_kafka_test_fatal_error() (by @ckb42)
- Ensure CMake sets disabled defines to zero on Windows (@benesch)
- librdkafka's build tooling now requires Python 3.x (the python3 interpreter).
Note: there was no v1.4.1 librdkafka release
Checksums
Release asset checksums:
* v1.4.2.zip SHA256 ac50da08be69365988bad3d0c46cd87eced9381509d80d3d0b4b50b2fe9b9fa9
* v1.4.2.tar.gz SHA256 3b99a36c082a67ef6295eabd4fb3e32ab0bff7c6b0d397d6352697335f4e57eb
1.4.0 (2020-04-02)
librdkafka v1.4.0
v1.4.0 is a feature release:
- KIP-98: Transactional Producer API
- KIP-345: Static consumer group membership (by @rnpridgeon)
- KIP-511: Report client software name and version to broker
Transactional Producer API
librdkafka now has complete Exactly-Once-Semantics (EOS) functionality, supporting the idempotent producer (since v1.0.0), a transaction-aware consumer (since v1.2.0) and full producer transaction support (in this release). This enables developers to create Exactly-Once applications with Apache Kafka.
See the Transactions in Apache Kafka page for an introduction and check the librdkafka transactions example for a complete transactional application example.
Security fixes
Two security issues have been identified in the SASL SCRAM protocol handler:
* The client nonce, which is expected to be a random string, was a static string.
* If sasl.username
and sasl.password
contained characters that needed escaping, a buffer overflow and heap corruption would occur. This was protected, but too late, by an assertion.
Both of these issues are fixed in this release.
Enhancements
- Add FNV-1a partitioner (by @Manicben, #2724).
The new
fnv1a_random
partitioner is compatible with Sarama'sNewHashPartitioner
partition, easing transition from Sarama to librdkafka-based clients such as confluent-kafka-go. - Added
rd_kafka_error_t
/RdKafka::Error
complex error type which provides error attributes such as indicating if an error is retriable. - The builtin mock broker now supports balanced consumer groups.
- Support finding headers in nonstandard directories in CMake (@benesch)
- Improved static library bundles which can now contain most dependencies.
- Documentation, licenses, etc, is now installed by
make install
- Bump OpenSSL to v1.0.2u (when auto-building dependencies)
Fixes
General:
* Correct statistics names in docs (@TimWSpence, #2754)
* Wake up broker thread based on next request retry.
Prior to this fix the next wakeup could be delayed up to 1 second regardless of next retry.
* Treat SSL peer resets as usual Disconnects, making log.connection.close work
* Reset buffer corrid on connection close to honour ApiVers and Sasl request priorities (@xzxxzx401, #2666)
* Cleanup conf object if failing to creat producer or consumer (@fboranek)
* Fix build of rdkafka_example project for windows, when using building it using Visual Studio 2017/2019 (by @Eliyahu-Machluf)
* Minor fix to rdkafka_example usage: add lz4 and zstd compression codec to usage (by @Eliyahu-Machluf)
* Let broker nodename updates propagate as ERR__TRANSPORT
rather than ERR__NODE_UPDATE
to avoid an extra error code for the application to handle.
* Fix erroneous refcount assert in enq_once_del_source (e.g., on admin operation timeout)
* Producers could get stuck in INIT state after a disconnect until a to-be-retried request timed out or the connection was needed for other purposes (metadata discovery, etc), this is now fixed.
Producer:
* flush()
now works with RD_KAFKA_EVENT_DR
* Fix race condition when finding EOS-supporting broker
Consumer:
* Consumers could get stuck after rebalance if assignment was empty
* Enforce session.timeout.ms
in the consumer itself (#2631)
* max.poll.interval.ms
is now only enforced when using subscribe()
* Fix consumer_lag
calculation for transactional topics
* Show fetch/no-fetch reason in topic
debugging
* Properly propagate commit errors per partition
* Don't send heartbeats after max.poll.interval.ms
is exceeded.
* Honour array size in rd_kafka_event_message_array()
to avoid overflow (#2773)
Checksums
Release asset checksums:
* v1.4.0.zip SHA256 eaf954e3b8a2ed98360b2c76f55048ee911964de8aefd8a9e1133418ec9f48dd
* v1.4.0.tar.gz SHA256 ae27ea3f3d0d32d29004e7f709efbba2666c5383a107cc45b3a1949486b2eb84
1.3.0 (2019-12-03)
librdkafka v1.3.0 release
This is a feature release adding support for KIP-392 Fetch from follower, allowing a consumer to fetch messages from the closest replica to increase throughput and reduce cost.
Features
- KIP-392 - Fetch messages from closest replica / follower (by @mhowlett)
- Added experimental (subject to change or removal) mock broker to make application and librdkafka development testing easier.
Fixes
- Fix
consumer_lag
in stats when consuming from broker versions <0.11.0.0 (regression in librdkafka v1.2.0).
Checksums
Release asset checksums:
* v1.3.0.zip SHA256 bd3373c462c250ecebea9043fb94597a11bd6e0871d3cde19019433d3f74a99e
* v1.3.0.tar.gz SHA256 465cab533ebc5b9ca8d97c90ab69e0093460665ebaf38623209cf343653c76d2
1.2.2 (2019-11-12)
librdkafka v1.2.2 release
v1.2.2 fixes the producer performance regression introduced in v1.2.1 which may affect high-throughput producer applications.
Fixes
- Fix producer insert msgq regression in v1.2.1 (#2450).
- Upgrade builtin lz4 to 1.9.2 (CVE-2019-17543, #2598).
- Don't trigger error when broker hostname changes (#2591).
- Less strict message.max.bytes check for individual messages (#993).
- Don't call timespec_get() on OSX (since it was removed in recent XCode) by @maparent .
- configure: add --runstatedir for compatibility with autoconf.
- LZ4 is available from ProduceRequest 0, not 3 (fixes assert in #2480).
- Address 12 code issues identified by Coverity static code analysis.
Enhancements
- Add warnings for inconsistent security configuration.
- Optimizations to hdr histogram (stats) rollover.
- Reorganized examples and added a cleaner consumer example, added minimal C++ producer example.
- Print compression type per message-set when
debug=msg
Checksums
Release asset checksums:
* v1.2.2.zip SHA256 7557b37e5133ed4c9b0cbbc3fd721c51be8e934d350d298bd050fcfbc738e551
* v1.2.2.tar.gz SHA256 c5d6eb6ce080431f2996ee7e8e1f4b8f6c61455a1011b922e325e28e88d01b53
1.2.1 (2019-10-09)
librdkafka v1.2.1 release
Warning: v1.2.1 has a producer performance regression which may affect high-throughput producer applications. We recommend such users to upgrade to v1.3.0
v1.2.1 is a maintenance release:
- Properly handle new Kafka-framed SASL GSSAPI frame semantics on Windows (#2542). This bug was introduced in v1.2.0 and broke GSSAPI authentication on Windows.
- Fix msgq (re)insertion code to avoid O(N^2) insert sort operations on retry (#2508) The msgq insert code now properly handles interleaved and overlapping message range inserts, which may occur during Producer retries for high-throughput applications.
- configure: added
--disable-c11threads
to avoid using libc-provided C11 threads. - configure: added more autoconf compatibility options to ignore
Checksums
Release asset checksums:
* v1.2.1.zip SHA256 8b5e95318b190f40cbcd4a86d6a59dbe57b54a920d8fdf64d9c850bdf05002ca
* v1.2.1.tar.gz SHA256 f6be27772babfdacbbf2e4c5432ea46c57ef5b7d82e52a81b885e7b804781fd6
1.2.0 (2019-09-19)
librdkafka v1.2.0 release
WARNING: There is an issue with SASL GSSAPI authentication on Windows with this release. Upgrade directly to v1.2.1 which fixes the issue.
v1.2.0 is a feature release making the consumer transaction aware.
- Transaction aware consumer (
isolation.level=read_committed
) implemented by @mhowlett. - Sub-millisecond buffering (
linger.ms
) on the producer. - Improved authentication errors (KIP-152)
Consumer-side transaction support
This release adds consumer-side support for transactions.
In previous releases, the consumer always delivered all messages to the application, even those in aborted or not yet committed transactions. In this release, the consumer will by default skip messages in aborted transactions.
This is controlled through the new isolation.level
configuration property which
defaults to read_committed
(only read committed messages, filter out aborted and not-yet committed transactions), to consume all messages, including for aborted transactions, you may set this property to read_uncommitted
to get the behaviour of previous releases.
For consumers in read_committed
mode, the end of a partition is now defined to be the offset of the last message of a successfully committed transaction (referred to as the 'Last Stable Offset').
For non-transactional messages there is no change from previous releases, they will always be read, but a consumer will not advance into a not yet committed transaction on the partition.
Upgrade considerations
linger.ms
default was changed from 0 to 0.5 ms to promote some level of batching even with default settings.
New configuration properties
- Consumer property
isolation.level=read_committed
ensures the consumer will only read messages from successfully committed producer transactions. Default isread_committed
. To get the previous behaviour, set the property toread_uncommitted
, which will read all messages produced to a topic, regardless if the message was part of an aborted or not yet committed transaction.
Enhancements
- Offset commit metadata (arbitrary application-specified data) is now returned by
rd_kafka_committed()
andrd_kafka_offsets_for_times()
(@damour, #2393) - C++: Added
Conf::c_ptr*()
to retrieve the underlying C config object. - Added
on_thread_start()
andon_thread_exit()
interceptors. - Increase
queue.buffering.max.kbytes
max to INT_MAX. - Optimize varint decoding, increasing consume performance by ~15%.
Fixes
General:
* Rate limit IO-based queue wakeups to linger.ms
, this reduces CPU load and lock contention for high throughput producer applications. (#2509)
* Reduce memory allocations done by rd_kafka_topic_partition_list_new()
.
* Fix socket recv error handling on MSVC (by Jinsu Lee).
* Avoid 1s stalls in some scenarios when broker wakeup-fd is triggered.
* SSL: Use only hostname (not port) when valid broker hostname (by Hunter Jacksson)
* SSL: Ignore OpenSSL cert verification results if enable.ssl.certificate.verification=false
(@salisbury-espinosa, #2433)
* rdkafka_example_cpp: fix metadata listing mode (@njzcx)
* SASL Kerberos/GSSAPI: don't treat kinit ECHILD errors as errors (@hannip, #2421)
* Fix compare overflows (#2443)
* configure: Add option to disable automagic dependency on zstd (by Thomas Deutschmann)
* Documentation updates and fixes by Cedric Cellier and @ngrandem
* Set thread name on MacOS X (by Nikhil Benesch)
* C++: Fix memory leak in Headers
(by Vladimir Sakharuk)
* Fix UBSan (undefined behaviour errors) (@PlacidBox, #2417)
* CONFIGURATION.md: escape ||
inside markdown table (@mhowlett)
* Refresh broker list metadata even if no topics to refresh (#2476)
Consumer:
* Make rd_kafka_pause|resume_partitions()
synchronous, making sure that a subsequent consumer_poll()
will not return messages for the paused partitions (#2455).
* Fix incorrect toppar destroy in OffsetRequest (@binary85, #2379)
* Fix message version 1 offset calculation (by Martin Ivanov)
* Defer commit in transport error to avoid consumer_close hang.
Producer:
* Messages were not timed out for leader-less partitions (.NET issue #1027).
* Improve message timeout granularity to millisecond precision (the smallest ffective message timeout will still be 1000ms).
* message.timeout.ms=0
is now accepted even if linger.ms
> 0 (by Jeff Snyder)
* Don't track max.poll.interval.ms
unless in Consumer mode, this saves quite a few memory barries for high-performance Producers.
* Optimization: avoid atomic fatal error code check when idempotence is disabled.
Checksums
Release asset checksums:
* v1.2.0.zip SHA256 6e57f09c28e9a65abb886b84ff638b2562b8ad71572de15cf58578f3f9bc45ec
* v1.2.0.tar.gz SHA256 eedde1c96104e4ac2d22a4230e34f35dd60d53976ae2563e3dd7c27190a96859
1.1.0 (2019-07-05)
librdkafka v1.1.0 release
v1.1.0 is a security-focused feature release:
- SASL OAUTHBEARER support (by @rondagostino at StateStreet)
- In-memory SSL certificates (PEM, DER, PKCS#12) support (by @noahdav at Microsoft)
- Pluggable broker SSL certificate verification callback (by @noahdav at Microsoft)
- Use Windows Root/CA SSL Certificate Store (by @noahdav at Microsoft)
ssl.endpoint.identification.algorithm=https
(off by default) to validate the broker hostname matches the certificate. Requires OpenSSL >= 1.0.2.- Improved GSSAPI/Kerberos ticket refresh
Upgrade considerations
- Windows SSL users will no longer need to specify a CA certificate file/directory (
ssl.ca.location
), librdkafka will load the CA certs by default from the Windows Root Certificate Store. - SSL peer (broker) certificate verification is now enabled by default (disable with
enable.ssl.certificate.verification=false
) %{broker.name}
is no longer supported insasl.kerberos.kinit.cmd
since kinit refresh is no longer executed per broker, but per client instance.
SSL
New configuration properties:
ssl.key.pem
- client's private key as a string in PEM formatssl.certificate.pem
- client's public key as a string in PEM formatenable.ssl.certificate.verification
- enable(default)/disable OpenSSL's builtin broker certificate verification.enable.ssl.endpoint.identification.algorithm
- to verify the broker's hostname with its certificate (disabled by default).- Add new
rd_kafka_conf_set_ssl_cert()
to pass PKCS#12, DER or PEM certs in (binary) memory form to the configuration object. - The private key data is now securely cleared from memory after last use.
Enhancements
- configure: Improve library checking
- Added
rd_kafka_conf()
to retrieve the client's configuration object - Bump
message.timeout.ms
max value from 15 minutes to 24 days (@sarkanyi, workaround for #2015)
Fixes
- SASL GSSAPI/Kerberos: Don't run kinit refresh for each broker, just per client instance.
- SASL GSSAPI/Kerberos: Changed
sasl.kerberos.kinit.cmd
to first attempt ticket refresh, then acquire. - SASL: Proper locking on broker name acquisition.
- Consumer:
max.poll.interval.ms
now correctly handles blocking poll calls, allowing a longer poll timeout than the max poll interval. - configure: Fix libzstd static lib detection
- rdkafka_performance: Fix for Misleading "All messages delivered!" message (@solar_coder)
- Windows build and CMake fixes (@myd7349)
Checksums
Release asset checksums:
* v1.1.0.zip SHA256 70279676ed863c984f9e088db124ac84a080e644c38d4d239f9ebd3e3c405e84
* v1.1.0.tar.gz SHA256 123b47404c16bcde194b4bd1221c21fdce832ad12912bd8074f88f64b2b86f2b