Repository navigation
Conversation
- Every range query also fits 6h and 24h windows (Alibaba and Google). - fetch_data.sh fetches the first day of Alibaba CallGraph and MCRRTUpdate (480 shards; was 6 hours), so a 24h window fits; MSMetrics and NodeMetrics were already a day. About 207 GB in all. - resolve_boundaries stitches instant samples in chunks of 60 steps. With 3-minute files almost every sample is a boundary one, and a day of MSRTMCR peaked at 220 GB in one pass; a sample counts for at most lookback steps, so each chunk reads its steps plus the lookback and the aggregates are the same (test). Peak is now 98 GB. - merge_samples draws its shares multinomially when the total reaches 1e9, numpy's limit for multivariate_hypergeometric (a day of Alibaba rt values); drawing 1e5 of 1e9 or more, the two agree (test). - results/skew_summary.csv refit: Google's 48 earlier rows are unchanged; Alibaba's change with 4x the evaluations per window. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FhcWJExZcmVqS6r6rEtjis
… fit NodeMetricsUpdate's first day spans steps 1..1439, one short of a 24h window, so node_cpu_by_nodeid and node_cpu_p99 had no 24h evaluation. Three 12-hour shards (steps 1..2159) give each 720; NodeMetrics is refit. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FhcWJExZcmVqS6r6rEtjis
zzylol
added a commit
that referenced
this pull request
Oct 8, 2026
- §4: plans priced by use, w_cpu * AUC(CPU) + w_mem * AUC(memory), CPU
elastic: ingest on ceil(rho) workers split by sample, compaction at
window close, query jobs; memory as ingest, storage, compaction and
query, each counted once; latency is a batch's longest chain (model in
sketch-bench #188). A peak-billed cost model was dropped.
- §5: latency is reported, in two versions: no SLA, with each method's
cost-latency frontier, and a batch latency SLA over {100 ms .. 10 s}.
- §3: PerQuery is the same MILP over each RQE's own candidates (no sharing).
- §6: the evaluated set is the mixed template set, with shared r in {1, 8}
and metrics m in {1, 8, 16}; the dashboard set is not evaluated.
- §7: AutoSketch's planning time is search plus its measured benchmark
(approxbench accuracy runs at 1e8 items); figures are the frontier, cost
vs. SLA and planning time; absolute costs only.
- §8-§10: PRs (#188, ASAPQuery #812, #138 stacked on #188), decisions Q3,
Q5-Q7, limitations (elastic CPU, compaction merge cost).
- §11: results on the synthetic mixed set.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FhcWJExZcmVqS6r6rEtjis
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
The AutoSketch vs. planner evaluation (#777) needs trace RQEs with 6h and 24h lookbacks. This PR refits the dataset analysis with them.
queries/alibaba_v2022.yamlandqueries/google_2011.yamlalso fits6hand24h.fetch_data.shfetches the first day of Alibaba CallGraph and MCRRTUpdate (480 three-minute shards; was 120, i.e. 6 hours), so a 24h window fits. MSMetrics and NodeMetrics already covered a day; Google already had about 7 days. About 207 GB in all.results/skew_summary.csvrefit:mcr_by_msname1h, n_evals 301 → 1381, lower θ 1.39 → 0.86.Two fixes the larger run needed
resolve_boundaries. With 3-minute files almost every sample is a file's first or last, and a day of MSRTMCR (3.9e9 rows) peaked at 220 GB on a 251 GB machine, then was killed. It now stitches in chunks of 60 steps. A sample counts for at mostlookbacksteps, so a chunk owning steps[start, end)reads the boundary samples up toend + lookbackand the aggregates are the same as one pass (test_boundary_chunks_match_one_pass). Peak RSS is now 98 GB.merge_samplesabove 1e9 values.numpy'smultivariate_hypergeometricrefuses a total of 1e9 or more, which a day of Alibabartvalues exceeds. At or above that limit each part's share is drawn multinomially, capped at its prefix. Drawing 1e5 of ≥ 1e9 values, with and without replacement agree (test_merge_samples_over_a_billion). Below the limit nothing changes.Run
On one 56-core, 251 GB CloudLab node each, 48 workers:
Checks
🤖 Generated with Claude Code
https://claude.ai/code/session_01FhcWJExZcmVqS6r6rEtjis