[
https://issues.apache.org/jira/browse/CASSANDRA-9673?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14708398#comment-14708398
]
Aleksey Yeschenko commented on CASSANDRA-9673:
----------------------------------------------
Rebased against most recent trunk and made some minor changes, here:
https://github.com/iamaleksey/cassandra/tree/9673-3.0. Didn't go very far.
Removed {{BATCHLOG_MUTATION}} Stage that Carl and I think is no longer needed
(but we'll bring it back if that's a mistake), and did some renames and minor
cleanup of unused methods.
Good work. There is only one real issue so far, with that fixed, we could
commit as is, but I have a bunch of nits that'd be nice to address: we are not
properly handling old-format batchlog mutations sent live from 2.1/2.2 nodes,
during live upgrades.
You need to modify {{MutationVerbHandler}} and add special treatment to
mutations for the old batchlog table.
Nits:
1. I would prefer {{BatchStoreMessage}} and {{BatchRemoveMessage}} to be dumb
and only be concerned about serialization/deserialization. This means no
{{getMutation()}} in either. For {{BatchRemoveMessage}}, the actual logic
should be in {{BatchRemoveVerbHandler}}.
2. I would like to separate {{Batch}} and {{BatchStoreMessage}} and put actual
saving to batchlog logic into a new {{Batch}} class - to be used both for local
(MV) writes and proper BL remote writes.
3. If moved to {{Batch}} (and if not, too), I would prefer the mutation
collections in them to only be set in the constructor.
4. Somehow we are calling {{LegacyBatchMigrator.convertBatchEntries()}} on
every replay, reverting Branimir's changes, instead of explicitly only doing it
once, upon first replay. I'm presuming that this is done as a workaround for
the live upgrade issue. I would prefer that we moved the migration to
{{CassandraDaemon.setup()}}, where legacy schema and hints migration already
get called.
> Improve batchlog write path
> ---------------------------
>
> Key: CASSANDRA-9673
> URL: https://issues.apache.org/jira/browse/CASSANDRA-9673
> Project: Cassandra
> Issue Type: Improvement
> Reporter: Aleksey Yeschenko
> Assignee: Stefania
> Labels: performance
> Fix For: 3.0 beta 2
>
> Attachments: 9673_001.tar.gz, 9673_004.tar.gz,
> gc_times_first_node_patched_004.png, gc_times_first_node_trunk_004.png
>
>
> Currently we allocate an on-heap {{ByteBuffer}} to serialize the batched
> mutations into, before sending it to a distant node, generating unnecessary
> garbage (potentially a lot of it).
> With materialized views using the batchlog, it would be nice to optimise the
> write path:
> - introduce a new verb ({{Batch}})
> - introduce a new message ({{BatchMessage}}) that would encapsulate the
> mutations, expiration, and creation time (similar to {{HintMessage}} in
> CASSANDRA-6230)
> - have MS serialize it directly instead of relying on an intermediate buffer
> To avoid merely shifting the temp buffer to the receiving side(s) we should
> change the structure of the batchlog table to use a list or a map of
> individual mutations.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
