docs(submitqueue): [DO NOT REVIEW][DO NOT LAND] Tango conflict research

[sbalabanov]

Summary

DRAFT RESEARCH ONLY — DO NOT REVIEW; DO NOT LAND. Leave this PR in draft. Do not request reviewers, mark it ready for review, or enable automatic merging. This branch contains a Go evaluator and a documented Tango/SubmitQueue conflict-analysis proposal; it does not deploy an analyzer.

Report: doc/rfc/submitqueue/tango-conflict-analysis.md. The earlier HTML report (https://personal.uberinternal.com/sergeyb/submitqueue-tango-conflict-analysis/index.html) reflects the superseded fixed-N workload and has not been updated.

Why?

A Tango-backed conflict analyzer stores a target signature per batch and compares each new batch against every in-flight batch. This research estimates, for a cold controller that must fetch and decode all B signatures, at what number of in-flight batches each signature design stops working because of size, materialization time or latency. go-code runs fewer than 100 in-flight batches today; the goal is to pick a design that scales.

What?

• Captured the complete go-code Bazel graph (2.97M targets) and measured its footprint.
• Pulled the exact per-request affected-target histogram for the go queue from Hive (58,876 requests; median 42, mean 2,527, max 2.88M). One batch is modeled as one request.
• Benchmarked three designs on closure-shaped affected sets drawn from that histogram, at B = 100–10,000: TangoSnapshot (baseline: the Tango response with dependency metadata, raw or gzip), NameKey (sorted labels) and ID64 (sorted 64-bit IDs plus an ID→name dictionary for conflict relaxation). For each, it serializes, deserializes and checks one candidate. A per-target Monte Carlo model extends the results to B = 100,000 and reports the first B exceeding example latency and heap budgets, under object-store and KV point-read profiles.
• Result: TangoSnapshot breaks on a single tail request at any B. NameKey holds to low thousands. ID64 is limited by round trips rather than bytes, so the choice of store matters more than the encoding.
• Simulated signatures computed once against the newest base. Plain affected-set overlap misses conflicts. Adding new-edge destinations (the production SubmitQueue rule) reduces but does not remove misses. Inheriting the signatures of structural dependencies removes them, at a cost in false conflicts. The report documents this as a policy trade-off.
• Removed superseded artifacts, the posting-index and cold-scan modes, and the root-level summary file.

Test Plan

• Evaluator unit tests: go test ./tool/tangograph-eval and ./tool/bazel test //tool/tangograph-eval/....
• BUILD and dependency checks: make check-gazelle && make check-tidy.
• Re-run the benchmark and the simulation with the commands under "Reproduce" in the report. The benchmark needs the local, uncommitted 2.8 GB Bazel stream; the last run took 13 min 48 s and peaked at 35 GB RSS.
• No live Tango endpoint, SubmitQueue controller or storage backend was changed or tested.

Revert Plan

Close this draft PR without merging. If it is ever merged contrary to this instruction, revert the research commits; they add only documentation and a standalone evaluator, not service behavior.

Issue Links

None — exploratory research requested directly; no tracking issue was provided.

🤖 Generated with Claude Code

AI Verification

Validated at f223612 on Sep 30 22:58 UTC · 40 files analyzed · 3s

Validator	Status	Issues
web-coverage	not_applicable	0
android-lint	not_applicable	0
web-lint	not_applicable	0
web-unit	not_applicable	0
merge-conflict	not_applicable	0
arc-lint	not_applicable	0
app-validation	not_applicable	0
ios-lint	not_applicable	0
web-repocheck	not_applicable	0
diff-template	not_applicable	0
fix-disclosure	not_applicable	0
android-coverage	not_applicable	0
java-lint	not_applicable	0
ios-test	not_applicable	0
web-typecheck	not_applicable	0
go-gazelle	not_applicable	0
arc-unit	not_applicable	0
go-proto-lint	not_applicable	0
go-thrift-lint	not_applicable	0
java-coverage	not_applicable	0
uber-one	not_applicable	0
go-coverage	failed	0
artifacts	completed	0
ureview	completed	0
custom	not_applicable	0
go-lint	failed	0

0 issues detected

_{Skipped validators: claude · EngWiki}

Prior runs

Run at f223612 on Sep 30 22:57 UTC · 40 files · 3s · 0 issues detected

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}