Support overriding the clickhouse binary with a PR build

[alexey-milovidov]

What

Adds a clickhouse_pr environment variable to the benchmark launcher. When set to a ClickHouse/ClickHouse pull-request number, the clickhouse* systems benchmark that PR's CI release build instead of the latest stable release. Every other system ignores it.

clickhouse_pr=81944 system=clickhouse ./run-benchmark.sh

How

• clickhouse_pr is set on the operator side (run-benchmark.sh), threaded to the VM through cloud-init.sh.in, and consumed only by the clickhouse* install scripts.
• A new shared helper, lib/download-clickhouse, replaces the inline curl https://clickhouse.com/ | sh in every clickhouse*/install. With no clickhouse_pr set it keeps the historical stable-download behaviour; with it set it resolves and downloads the PR build.
• "The PR build" = the newest PR commit whose Build (<arch>_release) CI check finished successfully for the machine's architecture (amd/arm), then the raw self-extracting binary that build uploaded to S3:

  https://clickhouse-builds.s3.amazonaws.com/PRs/<PR>/<SHA>/build_<arch>_release/clickhouse
  

  Commits are walked newest-first, so a tip commit whose build is still running, failed, or was skipped falls back to the last commit that does have a finished build — i.e. "the last commit with a finished release build". Uses the GitHub API (honours an optional GITHUB_TOKEN); fails loudly with a clear message if no successful build exists.

Filtering results

The chosen PR is written to the log as a ClickHouse PR: line and parsed by sink.parser into a new results.clickhouse_pr column so PR-build runs can be filtered apart from normal results (empty on a normal run). It is intentionally left out of the published output JSON, so the website payload is unchanged.

Important

prepare-database.sql is the schema definition; the live sink needs a manual migration to pick up the new column. SQL is in the PR discussion / provided to the operator.

Scope

Touches the seven clickhouse* systems that download the universal binary: clickhouse, clickhouse-trash-schema, clickhouse-web, clickhouse-parquet{,-partitioned}, clickhouse-datalake{,-partitioned}. Leaves clickhouse-tencent (hardcoded PR build) and clickhouse-cloud (managed) alone.

Testing

• Resolution dry-run against a real PR (#81944) for both amd and arm produced the correct, known-good S3 URL scheme.
• Ran the actual helper end-to-end with wget stubbed: native arm and forced-x86_64→amd both resolve correctly; no-clickhouse_pr falls back to the stable installer; missing-build case exits 1 with a clear message; candidate ordering is head-first, newest→oldest, deduped.
• ClickHouse PR: extraction verified: 81944 → 81944, empty → ''.
• All touched scripts pass bash -n; cloud-init.sh renders and validates with and without the PR set; the sink.parser SELECT (18 cols) aligns with the sink.results table.

🤖 Generated with Claude Code