Repository navigation
The fanout block publishes the slowest item's share of the pass (#3502) - #3507
Merged
Merged
Conversation
erikdarlingdata
enabled auto-merge (squash)
September 17, 2026 12:45
…verdict dominance was being read as (#3502) get_collection_health's per-collector fanout block published dominance - slowest_ms * items / run_ms, the slowest item against the MEAN item - and the tool description told readers to take the width-versus-concentration decision on it: near 1.0 the cost is width, around 2.0 or above one database dominates. That reading is numerically wrong at width, because the ratio's ceiling is items: at 72 databases, even a dominance of 10 is one database holding 14% of a pass. The #3477 reporter's evidence proved it with two instances scoring 3.58 (72 items) and 4.20 (15 items) - near-neighbour scores the guidance calls the same shape, whose slowest items' shares are 4.97% and 27.98%, 5.6x apart and wanting opposite remedies. The block now carries slowest_share_pct - slowest_ms / run_ms, which is dominance / items - the one-division-away number the decision actually turns on: low share means width and bounded parallelism is the lever, high share means one database dominates and a per-database schedule override or a stagger reaches exactly it. dominance stays for continuity as the evenness ratio - it is not wrong, it is just not a verdict. The share is DERIVED from dominance (share = dominance / items) on both SKUs' row models rather than recomputed from the columns, so the two figures can never drift apart or describe different runs, and it is null exactly when the fanout block is null - never a 0% that would read as "perfectly even" on a run that reported nothing. No schema change: the V80 columns already carry everything the share needs. Both SKUs move together because BothSkusToolDescriptions_StayByteIdentical holds the two descriptions to one contract: the corrected guidance routes the decision through the share and carries the two-instance counter-example as the permanent warning label, pinned alongside the field's wiring in CollectionLogFanoutRollupStoreTests (the anonymous-object emission no behavioural test reaches - the #3010 lesson). The V80 rung test keeps its two-shape arithmetic and now separates them by share; the new counter-example test pins the wide instance, the issue's dominance-of-10 limiting case, the share-saturates-at-100 endpoint, and the derivation itself, so a future edit that recomputes either figure from different columns fails loudly. The store round-trip asserts the share off real V80 columns beside the dominance it already read. Credit: the derivation, the counter-example, and the "publish the number the decision turns on" framing are the #3477 reporter's, verbatim from their evidence comment.
erikdarlingdata
force-pushed
the
fix/3502-fanout-share
branch
from
September 17, 2026 13:55
a7f5e97 to
59d619d
Compare
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.
Closes #3502 — ceremonially; the watcher closes it at merge.
The metric was honest; the guidance read it as a verdict
fanout.dominanceis the slowest item measured against the mean item, and the tool description told readers to treat it as a concentration verdict — "near 1.0 = width, around 2.0 or above = one database dominates." The #3477 reporter's evidence broke that reading cleanly: their two instances score 3.58 and 4.20 — near-neighbours, both far past the old bar — while the slowest item's actual share of the pass sits 5.6x apart (4.97% at 72 items, 27.98% at 15). At width, dominance can't say whether one database is worth chasing; the number the remediation decision turns on is the slowest item's share of the whole pass, which was one division away and never published.What ships
slowest_share_pctjoins the fanout block on both SKUs — the block is an engine-shared twin whose descriptions and payloads are pinned byte-identical, so a one-sided fix was structurally impossible — computed as a derivation from dominance (dominance ÷ items), so the two figures agree by construction and null exactly together. Dominance stays: it still answers how even the fan-out is; it just can't say whether the worst database matters, and its doc now says so. The description routes the width-vs-concentration decision through the share, with the reporter's counter-example compacted into the doc as its permanent warning label, and the retired verdict sentence is pinned absent so it can't drift back in.Tests
The counter-example as a fact (both instances, the dominance-of-10-at-72-items limiting case, share = 100 when the slowest is the run, and a same-inputs derivation pin over every row); the two-shapes arm re-routed through the share; the no-fanout arm reporting nothing rather than zero; the description pins (share formula present in both SKUs, retired guidance absent, and the emission wiring pinned in both anonymous objects — the mutation-invisibility lesson). Plus share asserts on the existing live-store round-trip. Credit for the derivation, the counter-example, and the publish-the-number-the-decision-turns-on framing belongs to the #3477 reporter, verbatim in the CHANGELOG.
Tests cannot run on macOS; verified by full-solution build and static tracing. CI is the first execution, review-dark per the known window.