vtgate: fix union merging through reference-table alternates - #20755
GrahamCampbell wants to merge 1 commit into
Conversation
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
NeedsWebsiteDocsUpdate
cf59812 to
04337dd
Compare
04337dd to
85b309b
Compare
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
85b309b to
43a042d
Compare
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #20755 +/- ##
===========================================
- Coverage 69.67% 4.92% -64.75%
===========================================
Files 1614 60 -1554
Lines 216793 8448 -208345
===========================================
- Hits 151044 416 -150628
+ Misses 65749 8032 -57717
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
43a042d to
44b8c2b
Compare
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
4ec9eb2 to
a07ae0c
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a07ae0cf8f
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
a07ae0c to
83729bf
Compare
There was a problem hiding this comment.
🟡 Changes recommended
Alternate rewrites can multiply volatile evaluation and shard-sensitive expanded operators, producing incorrect UNION results.
Once you've addressed the issues Copilot identified, you can request another Copilot review.
Review details
- Files reviewed: 5/5 changed files
- Comments generated: 2
- Review effort level: Balanced
| _ = Visit(route.Source, func(op Operator) error { | ||
| if horizon, ok := op.(*Horizon); ok && !horizon.computesRowByRow(ctx) { |
There was a problem hiding this comment.
Confirmed. With a subquery anywhere in the statement pushOrExpandHorizon expands instead of pushing, the Aggregator (or Limit, Distinct, Ordering) is then pushed under the single-shard reference route on its own, and the horizon check saw nothing: select count(*) from main_2.sharded_sales_ref where exists (select 1) union all select 0 from user.sales_extra planned as one scatter route running select count(*) from sales as sharded_sales_ref where exists (select 1 from dual) union all select 0 from sales_extra on every shard, one count per shard instead of one total. The same happened for LIMIT, DISTINCT, GROUP BY and HAVING, and with the subquery in the other branch (... union all select 0 from user.sales_extra where exists (select 1)). Main merges these through the bare prebuilt alternate and drops the operators entirely (select 1 from sales as sharded_sales_ref union all ...), so the path stays owned here. computesRowByRow now also stops on an Aggregator, Window, Distinct, Ordering or Limit under the moved route, the operators expandHorizon produces, so these pairings stay a Concatenate with the reference branch on main_2, while plain-column branches with a subquery (select col1 from main_2.sharded_sales_ref where exists (select 1) union all select cola from user.sales_extra) still merge onto the scatter routing. New goldens pin the aggregate and limit forms.
| resolvable = false | ||
| return io.EOF | ||
| } | ||
| if alt.VTable.Type != vindexes.TypeReference && alt.VTable.Keyspace.Sharded { |
There was a problem hiding this comment.
Pre-existing, and not widened here. A reference branch has always been folded into the sharded side's scatter routing once both sides are in one keyspace: on main, select id from ref where rand() < 0.5 union select id from user plans to the identical Distinct over a scatter route with rand() < 0.5 inside the pushed union, and union_cases.json already pins that shape for select id from ref_with_source where id in (null) union select id from user. The cross-keyspace UNION ALL form, select id from source_of_ref where rand() < 0.5 union all select id from user, merges into the same scatter route on main too, except that the prebuilt alternate drops the rand() < 0.5 predicate altogether (select id from ref_with_source as source_of_ref union all select id from user); this PR keeps the predicate and lets the DISTINCT form, which panicked, reach the shape the ALL form already had. Nothing in the planner classifies expressions as volatile, so handling it in this merge would be new policy that has to start with the same-keyspace case, and I'd rather track that separately than in this PR.
There was a problem hiding this comment.
Correction: the union_cases.json case I cited has a none-routed reference branch, so it shows the merged shape but not a live branch running per shard; the live same-keyspace form is not pinned on main, and select id from ref where rand() < 0.5 union select id from user is the reproduction.
44cdd80 to
eb7224a
Compare
Signed-off-by: Graham Campbell <hello@gjcampbell.co.uk>
eb7224a to
ebfb020
Compare
Description
A UNION DISTINCT between a reference table and a table in the keyspace holding one of its copies panics with
VT13001: [BUG] did not expect this method to be called. When the two sides' keyspaces differ, the planner makes them mergeable by swapping in the alternate route prebuilt at route-creation time, but unions merge later, during horizon planning, and the prebuilt route still holds the unplanned bare table it was created with. As soon as the union pushes its result columns into that source — the weight-string columns DISTINCT needs — the bare table operator panics inTable.AddColumn. UNION ALL survives because it pushes nothing.The fix stops using the prebuilt alternate as the merge payload. A declined cross-keyspace pairing is retried by resolving every real table under the moved side to its copy in the other side's keyspace through the VSchema — the table itself, a ReferencedBy copy, or a reference source, so routing rules stay in effect — and pointing the planned operators at those copies in place, undoing the move when the retry still cannot merge. Predicates and projections pushed after route creation are preserved because the operators are kept, physical renames keep the original name as an alias so column references still resolve, and later column pushes land on operators the planner has analyzed. A side already composed of a merged union is not moved: the remaining phases no longer normalize the shape that produces, so composite pairings stay split for now. A branch carrying an aggregate, limit, distinct or ordering stays split from a sharded source, including when a subquery elsewhere in the statement expanded its horizon, because moving it onto the sharded routing would run that operator once per shard. Join and subquery merging keep the prebuilt-route mechanism: they merge during physical planning, when every route source is still a bare table, and that path is covered and correct.
Fourteen planner cases in reference_cases.json cover both merge orders, the differing-physical-name rewrite, predicate preservation through the rewrite, the sharded-source routing preservation, the composite pairings that stay split, a no-copy control, and pin the already-working UNION ALL path; the four single-table DISTINCT cases fail on main with the panic. The reference-copy cases are skip_e2e like the rest of that file's, since the e2e fixture does not define the copy tables. Recursive CTE merging shares the swap through prepareInputRoutes and likely has the same latent panic, left for a follow-up. This is part one of the fix for #20754, and since the panic exists in released versions it is probably worth backporting.
Related Issue(s)
Part one of #20754.
Checklist
Deployment Notes
UNION DISTINCT queries pairing a reference table with a table in a keyspace holding one of its copies previously failed to plan with an internal error; they now plan as a single merged route. No migrations or configuration changes.
AI Disclosure
This PR was written primarily by Fable.
graph LR subgraph json [JSON support] direction TB PR20625["#20625 (merged)<br>mysql/json: fix MarshalTo discarding accumulated output for nested blob and bit values"] PR20632["#20632<br>vtgate: preserve IN value lists in complex aggregate projections"] PR21205["#21205<br>evalengine, mysql/json: fix JSON and TIME equality in compiled and interpreted paths"] PR20682["#20682 (draft)<br>evalengine: support constant-folded JSON values as literals"] PR20683["#20683 (draft)<br>evalengine, sqlparser: MySQL comparison domains; nested BETWEEN parentheses"] PR20626["#20626 (draft)<br>vtgate: support cross-shard JSON_ARRAYAGG and JSON_OBJECTAGG"] PR20625 --> PR20682 PR21205 --> PR20682 PR20682 --> PR20683 PR20632 --> PR20626 PR20683 --> PR20626 end subgraph union [Union routing] direction TB PR20628["#20628<br>vtgate: track per-source copies of pushed join predicates so merges skip them"] PR20629["#20629 (merged)<br>vtgate: fix None routing handling when merging unions"] PR20630["#20630 (draft)<br>vtgate: merge unions on join-predicate-free routings when all sources agree"] PR20755["#20755<br>vtgate: fix union merging through reference-table alternates"] PR20756["#20756 (draft)<br>vtgate: merge empty reference branches through rewritten copies"] PR20628 --> PR20630 PR20629 --> PR20630 PR20630 --> PR20756 PR20755 --> PR20756 end subgraph cte [Recursive CTEs] direction TB PR20759["#20759<br>vtgate: expand qualified stars without JOIN USING coalescing"] PR20633["#20633 (draft)<br>vtgate: validate the declared column list length of CTEs and derived tables"] PR20631["#20631 (draft)<br>vtgate: bind recursive CTE column filters as arguments in unmerged term queries"] PR20759 --> PR20633 PR20633 --> PR20631 end json ~~~ union ~~~ cte