[
https://issues.apache.org/jira/browse/IGNITE-28858?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Anton Vinogradov updated IGNITE-28858:
--------------------------------------
Description:
Follow-up to IGNITE-28857 (CDC replication test classes split). Two cheap
changes in the CDC replication tests cut their execution time roughly in half
(measured locally on CdcIgniteToIgniteReplicationThinClientTxTest):
1. walForceArchiveTimeout: 5000 -> 1000 ms. CDC consumes only archived WAL
segments, so every replication wait in a test pays up to a 5 s segment-archive
latency, several times per test. With 1 s the testActivePassiveReplication*
subset went from 124.7 s to 78.3 s (-37%). The same value is also used by the
PostgreSQL replication tests.
2. testWithExpiryPolicy TTL: 30 -> 10 s. The test waits for a real TTL
expiration twice (source and destination). With TTL 10 s (and 1 s
force-archive) the method went from 95.8 s (31.9 s per run) to 35.5 s (11.8 s
per run, -63%). 10 s still leaves an order-of-magnitude margin over the
observed replication lag.
Expected effect: ~25-30 min less CPU across the Cdc suite; the slowest Parallel
Tests batch after the split drops from ~10 to ~7 min.
was:
Follow-up to IGNITE-XXXXX (CDC replication test classes split). Two cheap
changes in the CDC replication tests cut their execution time roughly in half
(measured locally on CdcIgniteToIgniteReplicationThinClientTxTest):
1. walForceArchiveTimeout: 5000 -> 1000 ms. CDC consumes only archived WAL
segments, so every replication wait in a test pays up to a 5 s segment-archive
latency, several times per test. With 1 s the testActivePassiveReplication*
subset went from 124.7 s to 78.3 s (-37%). The same value is also used by the
PostgreSQL replication tests.
2. testWithExpiryPolicy TTL: 30 -> 10 s. The test waits for a real TTL
expiration twice (source and destination). With TTL 10 s (and 1 s
force-archive) the method went from 95.8 s (31.9 s per run) to 35.5 s (11.8 s
per run, -63%). 10 s still leaves an order-of-magnitude margin over the
observed replication lag.
Expected effect: ~25-30 min less CPU across the Cdc suite; the slowest Parallel
Tests batch after the split drops from ~10 to ~7 min.
> Reduce CDC replication tests execution time
> -------------------------------------------
>
> Key: IGNITE-28858
> URL: https://issues.apache.org/jira/browse/IGNITE-28858
> Project: Ignite
> Issue Type: Task
> Reporter: Anton Vinogradov
> Assignee: Anton Vinogradov
> Priority: Major
>
> Follow-up to IGNITE-28857 (CDC replication test classes split). Two cheap
> changes in the CDC replication tests cut their execution time roughly in half
> (measured locally on CdcIgniteToIgniteReplicationThinClientTxTest):
> 1. walForceArchiveTimeout: 5000 -> 1000 ms. CDC consumes only archived WAL
> segments, so every replication wait in a test pays up to a 5 s
> segment-archive latency, several times per test. With 1 s the
> testActivePassiveReplication* subset went from 124.7 s to 78.3 s (-37%). The
> same value is also used by the PostgreSQL replication tests.
> 2. testWithExpiryPolicy TTL: 30 -> 10 s. The test waits for a real TTL
> expiration twice (source and destination). With TTL 10 s (and 1 s
> force-archive) the method went from 95.8 s (31.9 s per run) to 35.5 s (11.8 s
> per run, -63%). 10 s still leaves an order-of-magnitude margin over the
> observed replication lag.
> Expected effect: ~25-30 min less CPU across the Cdc suite; the slowest
> Parallel Tests batch after the split drops from ~10 to ~7 min.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
