[
https://issues.apache.org/jira/browse/KNOX-3041?focusedWorklogId=1031638&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-1031638
]
ASF GitHub Bot logged work on KNOX-3041:
----------------------------------------
Author: ASF GitHub Bot
Created on: 22/Jul/26 09:20
Start Date: 22/Jul/26 09:20
Worklog Time Spent: 10m
Work Description: hanicz opened a new pull request, #1317:
URL: https://github.com/apache/knox/pull/1317
[KNOX-3041](https://issues.apache.org/jira/browse/KNOX-3041) - Eliminate
load-balancing race condition in HA dispatch
## What changes were proposed in this pull request?
Replace the peek-at-rewrite + shift-after-response backend selection with an
atomic pick-and-advance at dispatch time. The old pattern let concurrent
requests observe the same queue head for the entire HTTP round trip, so a slow
request would pin every concurrent request onto the same backend.
- Add URLManager#getActiveURLAndAdvance (and the HaProvider passthrough),
implemented as a synchronized poll+offer so concurrent callers see distinct
URLs whenever more than one is configured.
- setBackendUri now picks and rotates atomically; remove shiftActiveURL /
shiftCallback and the after-response shift in the sync and SSE dispatchers.
- Add updateBackendURL so a re-targeted request carries the picked backend's
base path (HA instances may publish per-server paths, e.g. HiveServer2),
recovering the source's base path by matching configured pool URLs.
**Two pre-existing bugs remain**
1. activeURL is set for the failed one after a failover for non LB-requests.
2. Failover might go to the same backend that just failed if there are
concurrent requests that rotate the URLs.
## How was this patch tested?
- Tested locally with custom service
- Tested locally with custom SSE service
- Unit tests
## Integration Tests
N/A
## UI changes
N/A
Issue Time Tracking
-------------------
Worklog Id: (was: 1031638)
Remaining Estimate: 0h
Time Spent: 10m
> Load Balancer backend selection synchronization issue
> -----------------------------------------------------
>
> Key: KNOX-3041
> URL: https://issues.apache.org/jira/browse/KNOX-3041
> Project: Apache Knox
> Issue Type: Bug
> Components: Server
> Reporter: Istvan Toth
> Priority: Major
> Time Spent: 10m
> Remaining Estimate: 0h
>
> I was trying to run a benchmark over Knox, with several Http client instances
> started at the same time on separate threads.
> I found that every single client was directed to the same backend, which was
> saved in the session, so instead of accessing each backend, every request was
> directed to a single one.
> While this was a benchmark/load test tool, any non-interactive client can
> exhibit similar behaviour, defating the point of load balancing.
> Looking at
> org.apache.knox.gateway.ha.dispatch.ConfigurableHADispatch.executeRequestWrapper(HttpUriRequest,
> HttpServletRequest, HttpServletResponse)
> , I think that the sitation could be improved by calling
> haProvider.makeNextActiveURLAvailable BEFORE the request is executed.
> It would be still racy, but the race window would be dramatically smaller,
> running some simple local code, which should be on the microsecond scale,
> instead of performing a full HTTP reqest, which is on on the millisecond
> scale at best.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
