[
https://issues.apache.org/jira/browse/CASSANDRA-12148?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15393714#comment-15393714
]
Joshua McKenzie commented on CASSANDRA-12148:
---------------------------------------------
I kept the index file around with the understanding that file contents can
reflect in kernel buffers before they're fsynced and thus be available to the
consumer before being durable. If the consumer relies on the reflected contents
and length of the file with a memory-mapped handle, for instance, they could
theoretically read ahead (in the case of high data production rates) and parse
CDC data that's never actually persisted to disk if the node goes down. At
least that's my understanding - correct me if I'm wrong.
As for continuing with a realtime parse from the last point in a segment: I
forgot that was explicitly mentioned on this ticket. It's a cleanly modular /
separate change to this one so I was planning on creating another ticket to
deal with that.
> Improve determinism of CDC data availability
> --------------------------------------------
>
> Key: CASSANDRA-12148
> URL: https://issues.apache.org/jira/browse/CASSANDRA-12148
> Project: Cassandra
> Issue Type: Improvement
> Reporter: Joshua McKenzie
> Assignee: Joshua McKenzie
>
> The latency with which CDC data becomes available has a known limitation due
> to our reliance on CommitLogSegments being discarded to have the data
> available in cdc_raw: if a slowly written table co-habitates a
> CommitLogSegment with CDC data, the CommitLogSegment won't be flushed until
> we hit either memory pressure on memtables or CommitLog limit pressure.
> Ultimately, this leaves a non-deterministic element to when data becomes
> available for CDC consumption unless a consumer parses live CommitLogSegments.
> To work around this limitation and make semi-realtime CDC consumption more
> friendly to end-users, I propose we extend CDC as follows:
> h6. High level:
> * Consumers parse hard links of active CommitLogSegments in cdc_raw instead
> of waiting for flush/discard and file move
> * C* stores an offset of the highest seen CDC mutation in a separate idx file
> per commit log segment in cdc_raw. Clients tail this index file, delta their
> local last parsed offset on change, and parse the corresponding commit log
> segment using their last parsed offset as min
> * C* flags that index file with an offset and DONE when the file is flushed
> so clients know when they can clean up
> h6. Details:
> * On creation of a CommitLogSegment, also hard-link the file in cdc_raw
> * On first write of a CDC-enabled mutation to a segment, we:
> ** Flag it as {{CDCState.CONTAINS}}
> ** Set a long tracking the {{CommitLogPosition}} of the 1st CDC-enabled
> mutation in the log
> ** Set a long in the CommitLogSegment tracking the offset of the end of the
> last written CDC mutation in the segment if higher than the previously known
> highest CDC offset
> * On subsequent writes to the segment, we update the offset of the highest
> known CDC data
> * On CommitLogSegment fsync, we write a file in cdc_raw as
> <segment_name>_cdc.idx containing the min offset and end offset fsynced to
> disk per file
> * On segment discard, if CDCState == {{CDCState.PERMITTED}}, delete both the
> segment in commitlog and in cdc_raw
> * On segment discard, if CDCState == {{CDCState.CONTAINS}}, delete the
> segment in commitlog and update the <segment_name>_cdc.idx file w/end offset
> and a DONE marker
> * On segment replay, store the highest end offset of seen CDC-enabled
> mutations from a segment and write that to <segment_name>_cdc.idx on
> completion of segment replay. This should bridge the potential correctness
> gap of a node writing to a segment and then dying before it can write the
> <segment_name>_cdc.idx file.
> This should allow clients to skip the beginning of a file to the 1st CDC
> mutation, track an offset of how far they've parsed, delta against the
> _cdc.idx file end offset, and use that as a determinant on when to parse new
> CDC data. Any existing clients written to the initial implementation of CDC
> need only add the <segment_name>_cdc.idx logic and checking for DONE marker
> to their code, so the burden on users to update to support this should be
> quite small for the benefit of having data available as soon as it's fsynced
> instead of at a non-deterministic time when potentially unrelated tables are
> flushed.
> Finally, we should look into extending the interface on CommitLogReader to be
> more friendly for realtime parsing, perhaps supporting taking a
> CommitLogDescriptor and RandomAccessReader and resuming readSection calls,
> assuming the reader is at the start of a SyncSegment. Would probably also
> need to rewind to the start of the segment before returning so subsequent
> calls would respect this contract. This would skip needing to deserialize the
> descriptor and all completed SyncSegments to get to the root of the desired
> segment for parsing.
> One alternative we discussed offline - instead of just storing the highest
> seen CDC offset, we could instead store an offset per CDC mutation
> (potentially delta encoded) in the idx file to allow clients to seek and only
> parse the mutations with CDC enabled. My hunch is that the performance delta
> from doing so wouldn't justify the complexity given the SyncSegment
> deserialization and seeking restrictions in the compressed and encrypted
> cases as mentioned above.
> The only complication I can think of with the above design is uncompressed
> mmapped CommitLogSegments on Windows being undeletable, but it'd be pretty
> simple to disallow configuration of CDC w/uncompressed CommitLog on that
> environment.
> And as a final note: while the above might sound involved, it really
> shouldn't be a big change from where we are with v1 of CDC from a C*
> complexity nor code perspective, or from a client implementation perspective.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
