bug: Ring and ledger count errors and coverage differently, cost reads slow down over time, and cost comments are stale

Cerrado
  • ¿Algo no está bien?
#1,566 0 comentarios 0 reacciones 0 asignados Ver en GitHub

Los mantenedores suelen responder en 1 día

Nadie ha tomado este issue todavía.

Evaluación

Dificultad
4/5
Tiempo estimado
3-5 días
Aptitud para principiantes
49/100
Tipo de issue
Error
Claridad
Bien especificado
Estado de actividad
Activo
Stack tecnológico
go
Área
backend, databases, performance

Línea de trabajo

Start with core/cost/usage/usage.go and core/cost/ledger/writer.go to compare the error and priceability rules, then inspect the archive read and write paths in core/session/archive and ledger queries in core/cost/ledger. Add or update focused tests for zero-usage successful responses, excluded errors, archive rotation and reads, and past-day caching. Done means the ledger and ring agree, archive reads avoid repeatedly decoding growing segments, and stale comments match the code.

Escrito por el modelo de indexación a partir del texto del issue.

Descripción

bug

Problem

The cost and session read paths have three problems. The ring (which serves the duration windows) and the ledger (which serves today/7d/month) count the same traffic by different rules. Two read paths get slower the longer the proxy runs. And several comments in core/cost describe code that has since changed. Line numbers are at ab5f01f26.

1. The ledger hides a coverage gap that the ring reports
  • Ledger (core/cost/ledger/writer.go:519-521): a request counts as priceable only if r.Model != "" && r.Tokens > 0.
  • Ring (core/cost/usage/usage.go:812-840): any successful response with a model counts as priceable, even with zero usage, and it is named in UnpricedBy.

Case: a truncated OpenAI-dialect stream returns 200, the model is set, and there is no usage block. inference-parser publishes no cost record for it. With 9 priced requests plus one of these:

  • window=today shows 9/9, which looks fully priced.
  • window=1h shows 9/10 and names the gap.

The ring's own comment (usage.go:821-825) calls the 9/9 result a bug it fixed. There is no ledger test for a successful response with zero usage.

2. The error rule is written twice, and only one copy was updated
  • countsAsError (usage.go:1279-1285) excludes ErrorKindUpgradeDeclined and ErrorKindDestinationPrivate. The latter was added in ab5f01f26.
  • The ledger (writer.go:514-516) still counts any StatusCode >= 400 || Denied.

So a parsed inference request refused with 403 destination_private counts as an error on today but not on 1h.

Fix for 1 and 2: export usage.CountsAsError and a single "priceable" check, and call both from the ledger writer.

3. Archive segments never rotate, and every read decodes from the start

core/session/archive/archive.go: a segment closes only after 60 minutes idle (:736), at the open-writer cap, or on delete, clear, rename or a write failure. There is no size or event-count limit.

  • planPrune never deletes an open segment (retention.go:54-63). A session busy all day therefore keeps one segment growing without limit, and maxBytes (2 GiB) can stay exceeded.
  • Page and Event decode the segment from its first event up to the cursor, and Earlier calls Page once per 500 events (read.go:44-81, 189-203). inference-router's wentTo calls Earlier on the request path with historyLimit = 2000. With 20k events in the segment, that is 4 page calls, each decompressing and decoding most of the segment, per request.

Fix: rotate in write after N events or M bytes, by calling closeWriter when s.w.info() passes the cap.

4. Every ledger query re-reads every day file in the window

core/cost/ledger/query.go:183-250 / store.go:493-530 open and fully decode each day file on every request, with no cache.

  • agentop polls today every minute and 7d/month every 5 minutes (cmd/agentop/tui/spend.go:211-217). On the 31st, month decodes 31 files per poll, and past days almost never change.
  • The endpoint has no auth.

Fix: cache decoded past days keyed by (path, size, mtime). Only today's file needs re-reading.

5. Comments that contradict the code
  • core/cost/pricing/cost.go:20-25 calls the accumulation wrap "an open defect", but Counts.Add saturates (usage.go:349-405).
  • cost.go:93-99 says MaxPlausibleTokens has "NO IMPORTER ON THIS BRANCH". The ledger writer imports it (writer.go:288-324).
  • cost.go:202-204 says plausibleTokenReport and RefusedTokenRequests don't exist. They do (usage.go:438, :122).
  • ledger/writer.go:929-933 says Dropped() has "No non-test caller today". sessionapi/usage.go:404 calls it.

These read like notes from the middle of a PR stack. A reader trusting them would conclude that sums can still wrap and that the token bounds are not enforced.

Fix: delete or correct them.

Found by a code audit of main at ab5f01f26. Items 3 and 4 were verified by reading the code; the decode cost was not measured.

Assisted-By: Claude (Anthropic AI) [email protected]

Lenguaje dominante
Go
Estrellas
21
Forks
43
Merge medio
15 h 38 min
PR fusionados (30 d)
234

Preparar el entorno

Primeros pasos

  1. Lee el issue completo y luego la guía de contribución del proyecto.
  2. Comenta en el issue que vas a ocuparte — evita que dos personas hagan lo mismo.
  3. Haz un fork del repositorio y trabaja en una rama.
  4. Abre un pull request que haga referencia al número del issue.

Más de rossoctl/cortex

Todos los issues de rossoctl/cortex

Issues similares

Más issues de Go

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.