[
https://issues.apache.org/jira/browse/KAFKA-16780?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Kamal Chandraprakash updated KAFKA-16780:
-----------------------------------------
Description:
h3. Logic to read aborted txns:
# When the consumer enables isolation_level as {{READ_COMMITTED}} and reads a
non-txn topic, then the broker has to
[traverse|https://github.com/apache/kafka/blob/trunk/core/src/main/scala/kafka/log/LocalLog.scala#L394]
all the local log segments to collect the aborted transactions since there
won't be any entry in the transaction index.
# The same
[logic|https://github.com/apache/kafka/blob/trunk/core/src/main/java/kafka/log/remote/RemoteLogManager.java#L1436]
is applied while reading from remote storage. In this case, when the FETCH
request is reading data from the first remote log segment, then it has to fetch
the transaction indexes of all the remaining remote-log segments, and then the
call lands to the local-log segments before responding to the FETCH request
which increases the time taken to serve the requests.
The [EoS Abort
Index|https://docs.google.com/document/d/1Rlqizmk7QCDe8qAnVW5e5X8rGvn6m2DCR3JR2yqwVjc]
design doc explains how the transaction index file filters out the aborted
transaction records.
The issue is when consumers are enabled with the {{READ_COMMITTED}} isolation
level but read the normal topics. If the topic is enabled with the transaction,
then we expect the transaction to either commit/rollback within 15 minutes
(default transaction.max.timeout.ms = 15 mins), possibly we may have to search
only a few remote log segments to collect the aborted txns.
was:
h3. Logic to read aborted txns:
# When the consumer enables isolation_level as {{READ_COMMITTED}} and reads a
non-txn topic, then the broker has to
[traverse|https://github.com/apache/kafka/blob/trunk/core/src/main/scala/kafka/log/LocalLog.scala#L394]
all the local log segments to collect the aborted transactions since there
won't be any entry in the transaction index.
# The same
[logic|https://github.com/apache/kafka/blob/trunk/core/src/main/java/kafka/log/remote/RemoteLogManager.java#L1436]
is applied while reading from remote storage. In this case, when the FETCH
request is reading data from the first remote log segment, then it has to fetch
the transaction indexes of all the remaining remote-log segments, and then the
call lands to the local-log segments before responding to the FETCH request
which increases the time taken to serve the requests.
The [EoS Abort
Index|https://docs.google.com/document/d/1Rlqizmk7QCDe8qAnVW5e5X8rGvn6m2DCR3JR2yqwVjc]
design doc explains how the transaction index file filters out the aborted
transaction records.
The issue is when consumers are enabled with the {{READ_COMMITTED}} isolation
level but read the normal topics. If the topic is enabled with the transaction,
then we expect the transaction to either commit/rollback within 15 minutes
(default transaction.max.timeout.ms = 15 mins), possibly we may have to search
only few remote log segments to collect the aborted txns.
> Txn consumer exerts pressure on remote storage when reading non-txn topic
> -------------------------------------------------------------------------
>
> Key: KAFKA-16780
> URL: https://issues.apache.org/jira/browse/KAFKA-16780
> Project: Kafka
> Issue Type: Task
> Reporter: Kamal Chandraprakash
> Priority: Major
>
> h3. Logic to read aborted txns:
> # When the consumer enables isolation_level as {{READ_COMMITTED}} and reads
> a non-txn topic, then the broker has to
> [traverse|https://github.com/apache/kafka/blob/trunk/core/src/main/scala/kafka/log/LocalLog.scala#L394]
> all the local log segments to collect the aborted transactions since there
> won't be any entry in the transaction index.
> # The same
> [logic|https://github.com/apache/kafka/blob/trunk/core/src/main/java/kafka/log/remote/RemoteLogManager.java#L1436]
> is applied while reading from remote storage. In this case, when the FETCH
> request is reading data from the first remote log segment, then it has to
> fetch the transaction indexes of all the remaining remote-log segments, and
> then the call lands to the local-log segments before responding to the FETCH
> request which increases the time taken to serve the requests.
> The [EoS Abort
> Index|https://docs.google.com/document/d/1Rlqizmk7QCDe8qAnVW5e5X8rGvn6m2DCR3JR2yqwVjc]
> design doc explains how the transaction index file filters out the aborted
> transaction records.
> The issue is when consumers are enabled with the {{READ_COMMITTED}} isolation
> level but read the normal topics. If the topic is enabled with the
> transaction, then we expect the transaction to either commit/rollback within
> 15 minutes (default transaction.max.timeout.ms = 15 mins), possibly we may
> have to search only a few remote log segments to collect the aborted txns.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
