Caleb Rackliffe created CASSANDRA-17523:
-------------------------------------------
Summary: Reduce histogram snapshot long[] allocation overhead
during speculative read and write threshold updates
Key: CASSANDRA-17523
URL: https://issues.apache.org/jira/browse/CASSANDRA-17523
Project: Cassandra
Issue Type: Improvement
Reporter: Caleb Rackliffe
Assignee: Caleb Rackliffe
Every 5 seconds with the default {{read_request_timeout}} (or in the old naming
scheme {{{}read_request_timeout_in_ms{}}}), a scheduled task updates the
speculation thresholds (for reads and writes) for all active tables. However,
there are a few issues with the way we do this:
1.) Whether or not the {{SpeculativeRetryPolicy}} implementations in use
actually looks at them, we create latency histogram snapshots to pass to
{{{}calculateThreshold(){}}}. We could trivially avoid this by having the
method take an argument of type {{Sampling}} and build the snapshot only when
necessary.
2.) The only reason we build the histogram snapshot is to find the new
threshold value for the percentile based policies.
{{EstimatedHistogramReservoirSnapshot}} creates copies of both the decaying and
non-decaying buckets, but we don’t use the non-decaying values at all for
percentile calculation. Just avoiding the non-decaying values array creation
would cut allocations in half.
Given even our snapshots aren’t perfectly consistent, it might also be possible
to calculate a percentile value directly from the reservoir’s decaying buckets,
although that might be less accurate, as new values could be added to the
buckets after a count is calculated.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
