[
https://issues.apache.org/jira/browse/KAFKA-4485?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Dong Lin updated KAFKA-4485:
----------------------------
Description:
As of current implementation, we will exclude follower from ISR if the begin
offset of FetchRequest from this follower is smaller than logEndOffset of
leader for more than replicaLagTimeMaxMs. Also, we will add a follower to ISR
if the beginOffset of FetchRequest from this follower is equal or larger than
high watermark of this partition.
This is problematic for the following reasons:
1) The criteria for ISR is inconsistent between maybeExpandIsr() and
maybeShrinkIsr(). Therefore a follower may be repeatedly remove and added to
the ISR (e.g. in the scenario described below).
2) A follower may be removed from the ISR even if its fetch rate can keep up
with produce rate. Suppose a produce keeps producing a lot of small requests at
high request rate but low byte rate (e.g. many mirror makers), and the follower
is always able to read all the available data at the time leader receives it.
However, the begin offset of fetch request will always be smaller than
logEndOffset of leader. Thus the follower will be removed from ISR after
replicaLagTimeMaxMs.
In the following we describe the solution to this problem.
Terminology:
- Definition of replica lag: we say a replica lags behind leader by X ms if its
current log end offset if equivalent to the log end offset of leader X ms ago.
- Definition of pseudo-ISR set: pseudo-ISR set of a partition = { replica |
replica belongs to the given partition AND replica's lag <= replicaLagTimeMaxMs}
- Definition of high-watermark of a partition: high-watermark of a partition is
the max(current high-watermark of the partition, min(offset of replicas in the
pseudo-ISR set of this partition))
- Definition of ISR set: ISR set of a partition = {replica | replica is in
pseudo-ISR set of the given partition AND log end offset of replica >=
high-watermark of the given partition}
Guarantee:
1) If a follower is close enough to the replica in the sense that its replica
lag <= replicaLagTimeMaxMs, then this follower will be in the pseudo-ISR set.
Thus the high-watermark will stop to increase until this follower's log end
offset >= high-watermark, at which moment this follower will be added to the
ISR set. This allows us the solve the 2nd problem described above.
2) If a follower lags behind leader for more than X ms, it will be removed out
of ISR set.
3) High watermark of a partition will never decrease.
4) For any replica in ISR set, its log end offset >= high-watermark.
Implementation:
1) For each follower, the leader keeps track of the time of the last fetch
request from this follower. Let's call it lastFetchTime. In addition, the
leader also maintains the log end offset of the leader at the lastFetchTime for
each follower. Let's call it lastFetchLeaderLEO. Both variables will be updated
after leader has processed a FetchRequest from a follower.
2) When leader receives FetchRequest from a follower, it will set the
follower's lag to lastFetchTime if begin offset of the FetchRequest >=
lastFetchLeaderLEO.
3) The leader can update pseudo-ISR set, high-watermark and ISR set of the
partition based on the lag of replicas of this partition, according to the
definition described above.
was:
As of current implementation, we will exclude follower from ISR if the begin
offset of FetchRequest from this follower is always smaller than logEndOffset
of leader for more than replicaLagTimeMaxMs.
Also, we will add a follower to ISR if the beginOffset of FetchRequest from
this follower is equal or larger than high watermark of this partition.
This is problematic for the following reasons:
1) The criteria for ISR is inconsistent between maybeExpandIsr() and
maybeShrinkIsr(). A follower may be repeatedly remove and added to the ISR
(e.g. in the scenario described below).
2) A follower may be removed from the ISR even if its fetch rate can keep up
with produce rate. Suppose a produce keeps producing a lot of small requests at
high request rate but low byte rate (e.g. many mirror makers), and the follower
is always able to read all the available data at the time leader receives it.
However, the begin offset of fetch request will always be smaller than
logEndOffset of leader. Thus the follower will be removed from ISR after
replicaLagTimeMaxMs.
The solution to the problem is the following:
A follower should be in ISR if begin offset of its FetchRequest >= max(high
watermark of partition, log end offset of leader at the time the leader
receives the previous FetchRequest). The follower should be removed from ISR if
this criteria is not met for more than replicaLagTimeMaxMs. Note that we are
comparing begin offset of FetchRequest with log end offset of leader at the
time the leader receives the previous FetchRequest as an approximate way to
compare the end offset of fetched data with log end offset of leader. This is
because we can not easily know the end offset of fetched data at the time
broker receives fetch request.
This solution makes the following guarantee:
1) If a follower is in ISR, then its log end offset >= high watermark of
partition at least sometime in the last replicaLagTimeMaxMs.
2) If a follower is not in ISR, then the end offset of its FetchRequest can not
catch up with log end offset of leader for more than replicaLagTimeMaxMs.
Either follower is in bootstrap phase, or the follower's average fetch rate is
smaller than average produce rate into the partition for the last
replicaLagTimeMaxMs.
> Follower should be in the isr if its FetchRequest has fetched up to the
> logEndOffset of leader
> ----------------------------------------------------------------------------------------------
>
> Key: KAFKA-4485
> URL: https://issues.apache.org/jira/browse/KAFKA-4485
> Project: Kafka
> Issue Type: Bug
> Affects Versions: 0.10.1.0
> Reporter: Dong Lin
> Assignee: Dong Lin
> Fix For: 0.10.2.0
>
>
> As of current implementation, we will exclude follower from ISR if the begin
> offset of FetchRequest from this follower is smaller than logEndOffset of
> leader for more than replicaLagTimeMaxMs. Also, we will add a follower to ISR
> if the beginOffset of FetchRequest from this follower is equal or larger than
> high watermark of this partition.
> This is problematic for the following reasons:
> 1) The criteria for ISR is inconsistent between maybeExpandIsr() and
> maybeShrinkIsr(). Therefore a follower may be repeatedly remove and added to
> the ISR (e.g. in the scenario described below).
> 2) A follower may be removed from the ISR even if its fetch rate can keep up
> with produce rate. Suppose a produce keeps producing a lot of small requests
> at high request rate but low byte rate (e.g. many mirror makers), and the
> follower is always able to read all the available data at the time leader
> receives it. However, the begin offset of fetch request will always be
> smaller than logEndOffset of leader. Thus the follower will be removed from
> ISR after replicaLagTimeMaxMs.
> In the following we describe the solution to this problem.
> Terminology:
> - Definition of replica lag: we say a replica lags behind leader by X ms if
> its current log end offset if equivalent to the log end offset of leader X ms
> ago.
> - Definition of pseudo-ISR set: pseudo-ISR set of a partition = { replica |
> replica belongs to the given partition AND replica's lag <=
> replicaLagTimeMaxMs}
> - Definition of high-watermark of a partition: high-watermark of a partition
> is the max(current high-watermark of the partition, min(offset of replicas in
> the pseudo-ISR set of this partition))
> - Definition of ISR set: ISR set of a partition = {replica | replica is in
> pseudo-ISR set of the given partition AND log end offset of replica >=
> high-watermark of the given partition}
> Guarantee:
> 1) If a follower is close enough to the replica in the sense that its replica
> lag <= replicaLagTimeMaxMs, then this follower will be in the pseudo-ISR set.
> Thus the high-watermark will stop to increase until this follower's log end
> offset >= high-watermark, at which moment this follower will be added to the
> ISR set. This allows us the solve the 2nd problem described above.
> 2) If a follower lags behind leader for more than X ms, it will be removed
> out of ISR set.
> 3) High watermark of a partition will never decrease.
> 4) For any replica in ISR set, its log end offset >= high-watermark.
> Implementation:
> 1) For each follower, the leader keeps track of the time of the last fetch
> request from this follower. Let's call it lastFetchTime. In addition, the
> leader also maintains the log end offset of the leader at the lastFetchTime
> for each follower. Let's call it lastFetchLeaderLEO. Both variables will be
> updated after leader has processed a FetchRequest from a follower.
> 2) When leader receives FetchRequest from a follower, it will set the
> follower's lag to lastFetchTime if begin offset of the FetchRequest >=
> lastFetchLeaderLEO.
> 3) The leader can update pseudo-ISR set, high-watermark and ISR set of the
> partition based on the lag of replicas of this partition, according to the
> definition described above.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
