[
https://issues.apache.org/jira/browse/CASSANDRA-15355?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
James Baker updated CASSANDRA-15355:
------------------------------------
Description:
In https://issues.apache.org/jira/browse/CASSANDRA-5025, pull based schema
updates were scheduled 1 minute after the schema change was first visible, so
as to prefer the push codepath as much as possible.
Unfortunately, this does not handle the case where there are many schema
changes happening - imagine a scenario where we create a table every 5 seconds
for 2 minutes - the first update tasks execute 60 seconds in and the schemas
may well be out of sync between nodes at that point.
In this case, there is some chance that when the task runs, the schemas are out
of sync because a subsequent schema update has occurred, and so the same
push/pull race has happened.
A fix is to modify the codepath such that the scheduled task is only run if the
other node's schema version is the same as when the task was scheduled. A
different (later scheduled) task should run otherwise.
For us, what we see is that when we have a reasonably large number of changes,
a few schema changes can have the unfortunate outcome of causing our nodes to
run out of memory and crash - if we have a 30 node cluster, create a table
every second for 2 minutes, and for some reason we pause for 10 seconds after
60 seconds with no progress, we can easily end up currently running 300 schema
pulls for a single node. These can cause further piling up which causes
cascading failures. This change stops that.
was:
In https://issues.apache.org/jira/browse/CASSANDRA-5025, pull based schema
updates were scheduled 1 minute after the schema change was first visible, so
as to prefer the push codepath as much as possible.
Unfortunately, this does not handle the case where there are many schema
changes happening - imagine a scenario where we create a table every 5 seconds
for 2 minutes - the first update tasks execute 60 seconds in and the schemas
may well be out of sync between nodes at that point.
In this case, there is some chance that when the task runs, the schemas are out
of sync because a subsequent schema update has occurred, and so the same
push/pull race has happened.
A fix is to modify the codepath such that the scheduled task is only run if the
other node's schema version is the same as when the task was scheduled. A
different (later scheduled) task should run otherwise.
For us, what we see is that when we have a reasonably large number of changes,
a few schema changes can have the unfortunate outcome of causing our nodes to
run out of memory and crash. This change stops that.
> Schema push/pull race on continuous schema changes
> --------------------------------------------------
>
> Key: CASSANDRA-15355
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15355
> Project: Cassandra
> Issue Type: Bug
> Reporter: James Baker
> Priority: Normal
> Labels: pull-request-available
> Time Spent: 10m
> Remaining Estimate: 0h
>
> In https://issues.apache.org/jira/browse/CASSANDRA-5025, pull based schema
> updates were scheduled 1 minute after the schema change was first visible, so
> as to prefer the push codepath as much as possible.
> Unfortunately, this does not handle the case where there are many schema
> changes happening - imagine a scenario where we create a table every 5
> seconds for 2 minutes - the first update tasks execute 60 seconds in and the
> schemas may well be out of sync between nodes at that point.
> In this case, there is some chance that when the task runs, the schemas are
> out of sync because a subsequent schema update has occurred, and so the same
> push/pull race has happened.
> A fix is to modify the codepath such that the scheduled task is only run if
> the other node's schema version is the same as when the task was scheduled. A
> different (later scheduled) task should run otherwise.
> For us, what we see is that when we have a reasonably large number of
> changes, a few schema changes can have the unfortunate outcome of causing our
> nodes to run out of memory and crash - if we have a 30 node cluster, create a
> table every second for 2 minutes, and for some reason we pause for 10 seconds
> after 60 seconds with no progress, we can easily end up currently running 300
> schema pulls for a single node. These can cause further piling up which
> causes cascading failures. This change stops that.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
