Online DDL: better error messages in cut-over phase #17052
Conversation
…ErrorFromError Signed-off-by: Shlomi Noach <[email protected]>
Signed-off-by: Shlomi Noach <[email protected]>
shlomi-noach
added
Type: Enhancement
shlomi-noach
requested review from
mattlord and
rohit-nayak-ps
as code owners
Review ChecklistHello reviewers! 👋 Please follow this checklist when reviewing this Pull Request. General
Tests
Documentation
New flags
If a workflow is added or modified:
Backward compatibility
|
vitess-bot
bot
added
NeedsBackportReason
shlomi-noach
removed
NeedsDescriptionUpdate
| @@ -896,7 +896,10 @@ func (e *Executor) cutOverVReplMigration(ctx context.Context, s *VReplStream, sh | |||
| defer cancel() | |||
| // Wait for target to reach the up-to-date pos | |||
| if err := tmClient.VReplicationWaitForPos(ctx, tablet.Tablet, s.id, replication.EncodePosition(pos)); err != nil { | |||
| return err | |||
| if s, _ := e.readVReplStream(ctx, s.workflow, true); s != nil { | |||
There was a problem hiding this comment.
As added bonus, why not also report what position we did land at.
| if merr, ok := err.(*sqlerror.SQLError); ok { | ||
| switch merr.Num { | ||
|
|
||
| if sqlErr, isSQLErr := sqlerror.NewSQLErrorFromError(err).(*sqlerror.SQLError); isSQLErr && sqlErr != nil { |
There was a problem hiding this comment.
This change is required because the underlying SQL error is now likely to be wrapped by a vitess error.
There was a problem hiding this comment.
Not sure if we have any wrapping, but I think places where we cast errors directly probably should use https://pkg.go.dev/errors#example-As these days. So if we're changing this, maybe also switch to that?
There was a problem hiding this comment.
sqlerror.NewSQLErrorFromError is even a bit more elaborate and can parse a MySQL error out of a string. We use that pretty consistently across the repo.
I'd say if we wanted to change that, it would be in a dedicated PR in one giant sweep.
There was a problem hiding this comment.
sqlerror.NewSQLErrorFromError
It's not about this function though? It's about the cast after?
There was a problem hiding this comment.
Ah, the cast at that same line... Rich story. We tried to get rid of it altogether, but hit the (weird, IMO) golang design decision where an interface (SQLError) can be non-nil where the underlying implementation is nil. So we kind of have to use the cast. What we know, though, is that the result of sqlerror.NewSQLErrorFromError is either a fully unwrapped sqlerror.SQLError or it is nil. So no need to use As on the result of this function.
Reference: #12574
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #17052 +/- ##
==========================================
+ Coverage 67.14% 67.15% +0.01%
==========================================
Files 1571 1571
Lines 252060 252064 +4
==========================================
+ Hits 169249 169286 +37
+ Misses 82811 82778 -33 ☔ View full report in Codecov by Sentry. |
Description
This PR adds more wrapping/clarifications to the
errorreturned bycutOverVReplMigration(). The function already logs the required information. However, the error returned by the function is also captured invitess_migrations.messagecolumn, and right nwo it is not informative enough. For example, it might read "due to context timeout" which isn't helpful if you don't know what specific operation was timing out. Again, the logs have the information, but it will be so much nicer to have more clarity in themessagecolumn.Related Issue(s)
None.
Checklist
Deployment Notes