[
https://issues.apache.org/jira/browse/CASSANDRA-13924?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Joseph Lynch updated CASSANDRA-13924:
-------------------------------------
Description:
I've been working on a way to keep data consistent without
scheduled/external/manual repair, because for large datasets repair is
extremely expensive. The basic gist is to introduce a new kind of hint that
keeps just the primary key of the mutation (indicating that PK needs repair)
and is recorded on replicas instead of coordinators during write time. Then a
periodic background task can issue read repairs to just the PKs that were
mutated. The initial performance degradation of this approach is non trivial,
but I believe that I can optimize it so that we are doing very little
additional work (see below in the design doc for some proposed optimizations).
My extremely rough proof of concept (uses a local table instead of HintStorage,
etc) so far is [in a
branch|https://github.com/apache/cassandra/compare/cassandra-3.11...jolynch:continuous_repair]
and has a rough [design
document|https://github.com/jolynch/cassandra/blob/continuous_repair/doc/source/architecture/continuous_repair.rst].
I'm working on getting benchmarks of the various optimizations, but I figured
I should start this ticket before I got too deep into it.
I believe this approach is particularly good for high read rate clusters
requiring consistent low latency, and for clusters that mutate a relatively
small proportion of their data (since you never have to read the whole dataset,
just what's being mutated). I view this as something that works _with_
incremental repair to reduce work required because with this technique we could
potentially flush repaired + unrepaired sstables directly from the memtable. I
also see this as something that would be enabled or disabled per table since it
is so use case specific (e.g. some tables don't need repair at all). I think
this is somewhat of a hybrid approach based on incremental repair, ticklers
(read all partitions @ ALL), mutation based repair (CASSANDRA-8911), and hinted
handoff. There are lots of tradeoffs, but I think it's worth talking about.
If anyone has feedback on the idea, I'd love to chat about it. [~bdeggleston],
[~aweisberg] I chatted with you guys a bit about this at NGCC; if you have time
I'd love to continue that conversation here.
was:
I've been working on a way to keep data consistent without
scheduled/external/manual repair, because for large datasets repair is
extremely expensive. The basic gist is to introduce a new kind of hint that
keeps just the primary key of the mutation (indicating that PK needs repair)
and is recorded on replicas instead of coordinators during write time. Then a
periodic background task can issue read repairs to just the PKs that were
mutated. The initial performance degradation of this approach is non trivial,
but I believe that I can optimize it so that we are doing very little
additional work (see below in the design doc for some proposed optimizations).
My extremely rough proof of concept (uses a local table instead of HintStorage,
etc) so far is [in a
branch|https://github.com/apache/cassandra/compare/cassandra-3.11...jolynch:continuous_repair]
and has a rough [design
document|https://github.com/jolynch/cassandra/blob/continuous_repair/doc/source/architecture/continuous_repair.rst.
I'm working on getting benchmarks of the various optimizations, but I figured
I should start this ticket before I got too deep into it.
I believe this approach is particularly good for high read rate clusters
requiring consistent low latency, and for clusters that mutate a relatively
small proportion of their data (since you never have to read the whole dataset,
just what's being mutated). I view this as something that works _with_
incremental repair to reduce work required because with this technique we could
potentially flush repaired + unrepaired sstables directly from the memtable. I
also see this as something that would be enabled or disabled per table since it
is so use case specific (e.g. some tables don't need repair at all). I think
this is somewhat of a hybrid approach based on incremental repair, ticklers
(read all partitions @ ALL), mutation based repair (CASSANDRA-8911), and hinted
handoff. There are lots of tradeoffs, but I think it's worth talking about.
If anyone has feedback on the idea, I'd love to chat about it. [~bdeggleston],
[~aweisberg] I chatted with you guys a bit about this at NGCC; if you have time
I'd love to continue that conversation here.
> Continuous/Infectious Repair
> ----------------------------
>
> Key: CASSANDRA-13924
> URL: https://issues.apache.org/jira/browse/CASSANDRA-13924
> Project: Cassandra
> Issue Type: Improvement
> Components: Repair
> Reporter: Joseph Lynch
> Priority: Minor
>
> I've been working on a way to keep data consistent without
> scheduled/external/manual repair, because for large datasets repair is
> extremely expensive. The basic gist is to introduce a new kind of hint that
> keeps just the primary key of the mutation (indicating that PK needs repair)
> and is recorded on replicas instead of coordinators during write time. Then a
> periodic background task can issue read repairs to just the PKs that were
> mutated. The initial performance degradation of this approach is non trivial,
> but I believe that I can optimize it so that we are doing very little
> additional work (see below in the design doc for some proposed optimizations).
> My extremely rough proof of concept (uses a local table instead of
> HintStorage, etc) so far is [in a
> branch|https://github.com/apache/cassandra/compare/cassandra-3.11...jolynch:continuous_repair]
> and has a rough [design
> document|https://github.com/jolynch/cassandra/blob/continuous_repair/doc/source/architecture/continuous_repair.rst].
> I'm working on getting benchmarks of the various optimizations, but I
> figured I should start this ticket before I got too deep into it.
> I believe this approach is particularly good for high read rate clusters
> requiring consistent low latency, and for clusters that mutate a relatively
> small proportion of their data (since you never have to read the whole
> dataset, just what's being mutated). I view this as something that works
> _with_ incremental repair to reduce work required because with this technique
> we could potentially flush repaired + unrepaired sstables directly from the
> memtable. I also see this as something that would be enabled or disabled per
> table since it is so use case specific (e.g. some tables don't need repair at
> all). I think this is somewhat of a hybrid approach based on incremental
> repair, ticklers (read all partitions @ ALL), mutation based repair
> (CASSANDRA-8911), and hinted handoff. There are lots of tradeoffs, but I
> think it's worth talking about.
> If anyone has feedback on the idea, I'd love to chat about it.
> [~bdeggleston], [~aweisberg] I chatted with you guys a bit about this at
> NGCC; if you have time I'd love to continue that conversation here.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
