Repository navigation
asap-tools: feed the sketchdb arm in the ClickHouse runner #874
Description
#868 feeds a sketch-bench workload table live into ClickHouse (dataset.paced_feeder in experiment_run_clickhouse.py), but refuses sketchdb mode. So nothing runs ASAPQuery's SQL arm on a live feed:
experiment_run_e2e.py(feat(asap-tools): run sketch-bench workload tables end to end with the paced feeder #841) feeds ASAPQuery, but its query client uses one protocol per server, so it can't send SQL query groups to ASAPQuery's SQL endpoint.experiment_run_clickhouse.py'ssketchdbmode sends SQL, but its engine ingestsdataset.local_data_file(ingest_json_config), which a fed config doesn't have.sketch_bench_workloads.pywrites<stem>/asapquery.sqlfor this arm, but no config reads it.
Needed
- Ingest: in a fed
sketchdbmode, start the query engine with remote-write ingest (noingest_json_config). - Shared feed: one
PacedFeederService.startwith two sinks, ASAPQuery's ingest (remote write, records as<metric>.<field>) and ClickHouse (which needs the data forforward_unsupported_queries), so both share one--start-msand seed. Thenset_data_startas for the other arms. - Generator: a
sketchdbarm config insketch_bench_workloads.pyusingasapquery.sqland a prebuilt planner output (--planner saved, feat(planner): --planner saved for SQL #866).
Open question
asapquery.sql has only the structured entries; PromQL entries have no ASAPQuery SQL. An arm with a subset of the queries has different query indices from baseline's, so results can't be matched by index. Either render the PromQL entries as ASAPQuery SQL too, or label results by query text/description instead of index.
Raised in review of #868.
🤖 Generated with Claude Code
Activity
milindsrivastava1997 commentedon Oct 15, 2026
CloudLab run of the fed sketchdb arm (#899–#901, plus #905): it works end to end.
Setup: one node (node1). sketch-bench
e2e-rate10multigroup table: 50 structured SQL queries over thehttpandflowsrecords metrics, fed at the table's rates (100k and 200k records/s). Saved plan:asap/cpu, cheapest (--planner saved --query-language sql). Starting delay 910 s, 600 s of queries (10 rounds, every 60 s).Results
destination=asap, HTTP 200). None were forwarded or unsupported.Every key matched across modes (0 missing, 0 extra). The p99 errors stay within DDSketch's 10% relative-error bound.
Caveat on
calculate_fidelity: it scores 34 of the 50 queries correctly (MAPE 1.0–1.4% cardinality, 4.9–6.1% p99). The other 16 come out as 0, because nothing is compared. They group byendpointafter a label that sorts after it (e.g.region, endpoint), and ASAPQuery's ClickHouse adapter writes label columns alphabetically while ClickHouse uses SELECT order. That's the limitation noted in #896. The table above uses an offline scorer that reads ASAPQuery's columns alphabetically. A fix would have the adapter emit SELECT order.Bugs found on the way (all fixed):
/clickhouse/queryendpoint: feat(asap-tools): sketchdb arm for sketch-bench ClickHouse workloads #901.🤖 Generated with Claude Code