[
https://issues.apache.org/jira/browse/BEAM-9039?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Robert Burke updated BEAM-9039:
---
Description:
Catch all task for any data channel stuckness issues, in particular if any
happen on errors. Re-open this issue if a new one is found.
The last known one I have is a race condition on DataChannel.readErr
Close off a race condition where a closing DataChannel might have new readers
created for it while it is failing, causing stuckness in the bundles.
In particular, the c.readErr must be interacted while c.mu is held.
Otherwise something like the following happens.
Given a channel C, and goroutines G1,G2.
# G1 A request for a new reader on C arrives, checks C.readErr finds it null.
# G2 An error occurs on reading. The lock is acquired, and C.readErr is set.
Readers are closed. The channel is officially closed with A.forceRecreate,
removing it from the DataManager cache.
# G1 calls A.makeReader, and creates a new reader there.
There could be an arbitrary number of G1s.
was:
Catch all task for any data channel stuckness issues, in particular if any
happen on errors.
The last known one I have is a race condition on DataChannel.readErr
Close off a race condition where a closing DataChannel might have new readers
created for it while it is failing, causing stuckness in the bundles.
In particular, the c.readErr must be interacted while c.mu is held.
Otherwise something like the following happens.
Given a channel C, and goroutines G1,G2.
# G1 A request for a new reader on C arrives, checks C.readErr finds it null.
# G2 An error occurs on reading. The lock is acquired, and C.readErr is set.
Readers are closed. The channel is officially closed with A.forceRecreate,
removing it from the DataManager cache.
# G1 calls A.makeReader, and creates a new reader there.
There could be an arbitrary number of G1s.
> Fix Datachannel stuckness on errors
> ---
>
> Key: BEAM-9039
> URL: https://issues.apache.org/jira/browse/BEAM-9039
> Project: Beam
> Issue Type: Sub-task
> Components: sdk-go
>Affects Versions: Not applicable
>Reporter: Robert Burke
>Assignee: Robert Burke
>Priority: Major
> Fix For: Not applicable
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> Catch all task for any data channel stuckness issues, in particular if any
> happen on errors. Re-open this issue if a new one is found.
> The last known one I have is a race condition on DataChannel.readErr
> Close off a race condition where a closing DataChannel might have new readers
> created for it while it is failing, causing stuckness in the bundles.
> In particular, the c.readErr must be interacted while c.mu is held.
> Otherwise something like the following happens.
> Given a channel C, and goroutines G1,G2.
> # G1 A request for a new reader on C arrives, checks C.readErr finds it null.
> # G2 An error occurs on reading. The lock is acquired, and C.readErr is set.
> Readers are closed. The channel is officially closed with A.forceRecreate,
> removing it from the DataManager cache.
> # G1 calls A.makeReader, and creates a new reader there.
> There could be an arbitrary number of G1s.
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
