[
https://issues.apache.org/jira/browse/FLINK-40776?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18118163#comment-18118163
]
Utkarsh Vashishtha commented on FLINK-40776:
--------------------------------------------
The fix is small and committed in the fork we run, I'll create a PR upstream if
the ticket can be assigned.
> PARSE_URL compiles the query-parameter pattern on every row
> -----------------------------------------------------------
>
> Key: FLINK-40776
> URL: https://issues.apache.org/jira/browse/FLINK-40776
> Project: Flink
> Issue Type: Improvement
> Components: Table SQL / Runtime
> Reporter: Utkarsh Vashishtha
> Priority: Major
>
> {{SqlFunctionUtils.parseUrl(urlStr, partToExtract, key)}} compiles a new
> {{Pattern}} on every invocation:
> {code:java}
> Pattern p = Pattern.compile("(&|^)" + Pattern.quote(key) + "=([^&]*)");
> {code}
> The pattern depends only on {{{}key{}}}. The URL parse immediately above
> already goes through {{{}URL_CACHE{}}}.
> h3. Fix
> Route the pattern through a {{ThreadLocalCache}} keyed on {{{}key{}}},
> matching {{URL_CACHE}} in the same class:
> {code:java}
> private static final ThreadLocalCache<String, Pattern>
> QUERY_PARAM_PATTERN_CACHE =
> ThreadLocalCache.of(k -> Pattern.compile("(&|^)" + Pattern.quote(k) +
> "=([^&]*)"));
> {code}
>
> This fix already runs in our Flink fork and performance benchmarks show an
> improvement ranging till 4x, depending on key being static (most common /
> highest improvement) v/s dynamic (dependent on cardinality).
> Found while investigating per-row regex costs in a production Flink SQL
> deployment.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
