Did the revision help?
Keep the original. Add the revision. See what the evidence supports.
Paste analysis links from your agent or Save & share. Compare locally, then save both runs and your notes in one encrypted seven-day link. Nothing uploads until you choose.
Reopen a saved comparison
Or import an offline copy
Ask your agent for a candidate
Keep the baseline here. Paste this into the conversation where you are revising the query, then paste the returned analysis link into Candidate below.
Keep our original ClickHouse query and its evidence unchanged as the baseline. For the revision we agreed on, follow https://chquery.com/llms-full.txt and return a separate candidate analysis link for CH Query, without uploading. Ask which baseline and revision if they are unclear. Use only permissions already explicitly granted in this conversation; this prompt grants no new database access, query execution, benchmarks, setting changes, uploads or sharing. Keep measurements tied to the exact query they describe. State whether result correctness was checked, and distinguish plan changes from measured improvements.
Hypothesis & correctness assertions
Author assertions only. Equal SQL, plans or row counts do not prove equivalent results.
Import a baseline and candidate bundle to begin.
Explore comparison evidence
Evidence pages preserve source text. Operator lists use role-local addresses, not cross-run identities. Correspondence is heuristic.
Save & share this comparison
One link reopens the baseline, candidate, and included hypothesis and correctness notes. Review the snapshot, then choose encrypted seven-day storage. No file handoff needed.
Partial redaction can leave identifiers, small numbers, DDL, provenance and unsupported fields sensitive. Inspect the complete outgoing snapshot.
Review your snapshot
Offline copies & reports
Optional fallback. Reviewed copies require the same review and confirmation above. Full source is unredacted and can contain private metadata.
See the shape of your query.
Paste ClickHouse EXPLAIN output. Analysis stays in your browser. You choose what to share.
Or hand this to your agent
Use https://chquery.com/llms.txt to review the ClickHouse query in this conversation. Send me one CH Query link.
Follow the repeated reads of actors.
From the analysis workspace: a 14-second look at the plan and its evidence.
- Overview
This public capture has 52 plan operators. The workspace groups them into 34 stages, preserving their dependencies.
- Repeated reads
Seven separate
ReadFromMergeTreeoperators refer todefault.actors. Each reports 4/4 selected granules in the plan, not measured runtime work. - Supporting evidence
EXPLAIN ESTIMATE supplies 235,753 rows, 7 parts and 28 marks for the table as a whole. Per-read row counts and runtime are unknown; no speedup is measured.
Public ClickHouse Playground capture · September 8, 2026 · 26.9.1.52. Inspect the captured bundle. Opening the example keeps your in-tab personal work and draft available.
Query workspace
Example · a cohort leaderboard over actors
Explore the plan Loading the example plan…
Query
More
Discuss a scoped review of one query
Inspect the evidence and missing context first. To discuss a review with Shenandoah, email Max Sumrall with the question you want to resolve. Decide what evidence to share after agreeing the scope. This opens your email app without attaching SQL, notes or a link.
No plan loaded
Open the wizard or import a bundle.
Repeated reads of actors
One thing to try — inspect whether the location aggregates could share work.
Findings.
Load a plan to see findings.
Server.
No server information in this bundle.