Skip to content

Commit 9cdbb73

Browse files
scuttletobyclaude
andauthored
feat(server): keep a history of provider usage-limit readings (#24)
* feat(server): keep a history of provider usage-limit readings Pacing needs to know how fast each account's meter moves, not just where it stands now. The server only keeps the last usage-limit snapshot per provider instance, so the burn rate, an early reset or another client's use cannot be seen. UsageReadingsHistory appends one JSON line per reading to <state>/usage-readings/<instanceId>.jsonl whenever an instance's meter moved (0.1 point), its account or availability changed, or 30 minutes passed unchanged. A republished or older snapshot is never recorded. At start it compacts every history: readings older than 35 days and lines torn by a crash are dropped, at most 20 000 readings are kept per instance, and a history with nothing left is removed. It is file-based on purpose: the fork carries no database migrations. Co-Authored-By: Claude Opus 5.5 <[email protected]> * Usage readings: account per series, safe reads, bounded and thinned history - A read error other than "file missing" no longer counts as an empty history: that instance is skipped and logged, its file is never compacted or removed, and readings() fails with UsageHistoryUnreadable. - Each reading records the account the Limits UI groups by (providerAccountKey, from accountKey), as a short sha256 hash. An account change starts a new series; readingsOfAccount groups instances. - Record rule matches its words: a window moved |delta| >= 0.1 point, a reset moved >= 1 minute, or availability, window set, kind or length changed; windows are compared by id. - Compaction interval, cap, retention, heartbeat and thinning are options. Readings older than 2 days thin to one per 15 minutes plus every structural change, so 35 days fit the cap. - Retention and cap are based on the newest reading, not the clock, so a clock far ahead at start cannot wipe the history. Readings stamped ahead of the clock are not served or used, but stay on disk. - A torn last line gets a newline before the next append. - Lines that no longer decode are skipped one by one and kept on disk. - Reads are served from memory after the first load. - Tests: 21, each new guard checked to fail its test when removed. Co-Authored-By: Claude Opus 5.5 <[email protected]> * Usage readings: keep the account on a failed probe, clock rule by append order - A probe that reports less of the account (no identity, or no workspace for the same login) keeps the instance's last known account, or the one workspace other instances of that login report in the snapshot, so a failed probe never looks like an account swap. Accounts are stored as hashed <base> or <base>.<workspace>. - A reading appended before an earlier-stamped one came from a clock that was ahead: readers ignore it and compaction drops it. Readings ahead of a clock that is behind, with no later append, stay. - A failing history (unreadable, appends failing) is logged once per instance until it works again. - Compaction removes a temp folder left by a crash during compaction. - Tests for reset appear/disappear, window kind, window replaced, a torn append in a running server, reads from memory and lock-free reads. Co-Authored-By: Claude Opus 5.5 <[email protected]> * Usage readings: no deletion on clock changes, segments, real-write recovery, no borrowed carry - Clock: drop the append-order rule. No reading is deleted or hidden because its stamp disagrees with another or with the clock; deletion is only retention, the cap and thinning, counted from the newest reading. Readers still skip readings stamped ahead of now. `readings` returns append order with a `segment` per reading that goes up where a stamp goes backwards; thinning starts again at a break. `newestMs` is the newest reached stamp. - Log "recording works again" only after a real write or a real read of a file that had failed, never after an update with nothing to write. - A workspace borrowed from another instance of the same login is marked `borrowed` and never carried as the instance's last known account; two workspaces of one login in a snapshot are never merged. Co-Authored-By: Claude Opus 5.5 <[email protected]> --------- Co-authored-by: Claude Opus 5.5 <[email protected]>
1 parent 954235f commit 9cdbb73

4 files changed

Lines changed: 1756 additions & 0 deletions

File tree

‎apps/server/src/server.ts‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -115,6 +115,7 @@ import * as ProviderInstanceRegistry from "./provider/ProviderInstanceRegistry.t
115115
import * as ProviderAdapterRegistry from "./orchestration-v2/ProviderAdapterRegistry.ts";
116116
import * as ProviderRegistry from "./provider/ProviderRegistry.ts";
117117
import * as ProviderUsageLimitsIngestion from "./provider/ProviderUsageLimitsIngestion.ts";
118+
import * as UsageReadingsHistory from "./usage/UsageReadingsHistory.ts";
118119
import * as UsageLimitSources from "./usage/UsageLimitSources.ts";
119120
import * as ProjectFaviconResolver from "./project/ProjectFaviconResolver.ts";
120121
import * as T3ProjectFileLoader from "./project/T3ProjectFileLoader.ts";
@@ -681,6 +682,7 @@ const layerRuntimeCoreDependenciesBase = Layer.mergeAll(
681682
}),
682683
).pipe(Layer.provide(DelegatedTaskQuestionReactor.layer), Layer.provide(ProjectionStoreV2.layer)),
683684
ProviderUsageLimitsIngestion.layer,
685+
UsageReadingsHistory.layer,
684686
layerProviderInstallationRefresh,
685687
ProviderInstallation.layer,
686688
ReplayMarkers.layer,

0 commit comments

Comments
 (0)