Roman Puchkovskiy created IGNITE-20668:
------------------------------------------
Summary: Increase wait after a DDL to account for idle safe-time
propagation period
Key: IGNITE-20668
URL: https://issues.apache.org/jira/browse/IGNITE-20668
Project: Ignite
Issue Type: Improvement
Reporter: Roman Puchkovskiy
Assignee: Roman Puchkovskiy
Fix For: 3.0.0-beta2
For an RO transaction to work correctly (that is, the same query returns the
same result), an RO transaction sometimes needs to wait. In most scenarios, if
the readTimestamp is 'now', it ruins the performance (as such transaction have
to wait almost always, up to idle safe-time propagation period). To avoid this,
readTimestamp is usually chosen in the past as 'now minus idle safe-time
propagation period'.
But if a DDL has just happened, then looking in the past might look past the
DDL, so a table might not be found, or a stale version of it might be seen by
the query.
The 'correct' way to solve this is described in IGNITE-20378, but it will take
some time to implement.
We need the RO-related optimization right now, so we could just increase the
after-DDL wait by the same idle safe-time propagation period as a workaround.
We should also make this period configurable to have ability to make it lower
in our tests (so that they don't be too slow waiting for a second after each
table creation).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
