[
https://issues.apache.org/jira/browse/IMPALA-15234?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18105826#comment-18105826
]
Aleksandr Efimov commented on IMPALA-15234:
-------------------------------------------
Working on this. To keep the first step small I went with a WebUI endpoint
rather than a statement: CacheBackend gets a clear(), InMemoryCacheBackend
calls invalidateAll(), and it is exposed as /clear_hbo_stats next to the
existing /resource_pool_reset, reachable through a new clearHboStats() JNI
entry point. That clears the cache of the coordinator serving the request,
which is what the e2e tests need since they talk to a single coordinator.
A statement like CLEAR HBO STATS would need parser and analyzer work, and
raises the question of whether it should reach every coordinator rather than
just the one handling the request. Happy to follow up with that if you'd prefer
it — it would call the same clear() underneath.
> Provide a way to clear HBO cache
> --------------------------------
>
> Key: IMPALA-15234
> URL: https://issues.apache.org/jira/browse/IMPALA-15234
> Project: IMPALA
> Issue Type: New Feature
> Components: Backend
> Reporter: Quanlong Huang
> Assignee: Aleksandr Efimov
> Priority: Major
>
> We should provide a command (in impala-shell) or WebUI endpoint to clear HBO
> stats, in case when "bad" stats are stored, to prevent poisoning future query
> plans.
> In e2e tests while testing expected HBO stats are stored, we also need this
> to clear the cache, to make sure they are not accidentially stored by
> previous queries.
> The HBO cache are keyed by hash strings so it's hard to identify which keys
> belong to a specific table. As a first step, we should at least support
> clearing the entire cache.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
