Amrit Sarkar created SOLR-18670:
-----------------------------------

             Summary: Async Collection API commands keep polling sub-requests 
that already failed and report a misleading notfound error
                 Key: SOLR-18670
                 URL: https://issues.apache.org/jira/browse/SOLR-18670
             Project: Solr
          Issue Type: Bug
          Components: SolrCloud
            Reporter: Amrit Sarkar


SOLR-18081 added a check to {{ShardRequestTracker.waitForAsyncCallsToComplete}} 
in {{CollectionHandlingUtils}} so it doesn't wait on async core admin calls 
that already failed:
{code:java}
Object failure =
    results._get("failure/" + requestKey(asyncCmdInfo.nodeName, 
asyncCmdInfo.coreNodeName));
if (failure != null) {
  return;
}
{code}
It has two problems:
* {{requestKey}} returns {{node/coreNodeName}}, and {{_get}} splits its path on 
{{/}}. The lookup goes to {{failure -> node -> coreNodeName}}, but 
{{addFailure}} stores the flat key {{node/coreNodeName}}. Every caller passes a 
core node name, so the check never matches.
* If it did match, {{return}} would stop waiting on every remaining async call, 
not just the failed one. It should be {{continue}}.

So today, when an async sub-request fails, the tracker still polls 
REQUESTSTATUS for it. If the core never registered the task, it answers 
{{notfound}}. After 5 retries a second apart, the command fails with "Invalid 
status request for requestId ... notfound". That hides the real failure, and 
the remaining sub-requests are never waited on. The same polling happened 
before SOLR-18081, so this isn't a regression.

Fix: look the key up in the {{failure}} map directly and {{continue}}.



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to