GeezerVison

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
docs	Skipped		May 7, 2026 6:50pm

PR Summary

Medium Risk
Changes billing write semantics by removing the atomic transaction between usage_log inserts and user_stats counter updates, allowing counters to drift when the update fails. This reduces pgbouncer connection pinning under high concurrency but risks temporary mismatches in reported totals/period costs.

Overview
Updates recordUsage to stop wrapping usage_log inserts and user_stats counter increments in a single DB transaction to alleviate pgbouncer pool contention under concurrent usage.

usage_log inserts remain required (errors propagate), while the user_stats update is now best-effort: missing-row and update failures no longer throw/rollback and instead log warnings (including whether entries existed and which additionalStats keys were requested).

^{Reviewed by Cursor Bugbot for commit ec5793f. Bugbot is set up for automated code reviews on this repo. Configure here.}

Greptile Summary

This PR removes the db.transaction() wrapper from recordUsage to fix pgbouncer pool exhaustion caused by long-lived row locks on user_stats under high per-user concurrency. The usage_log INSERT now commits immediately and propagates errors to callers, while the userStats UPDATE is demoted to best-effort with all failures logged and swallowed.

• Behavioral change: previously, a missing user_stats row caused the usage_log INSERT to roll back; it now commits the INSERT regardless, which is a strict improvement for data durability.
• Accepted trade-off: any user_stats UPDATE failure leaves currentPeriodCost / totalCost counters under-counted until reconciled from usage_log; billing-limit enforcement may allow slightly more spend than configured during drift windows.
• Catch-all error handling: the catch block treats transient and permanent failures identically, permanently dropping the counter increment rather than retrying."

Confidence Score: 4/5

Safe to merge — the fix correctly addresses pool contention and the data durability trade-offs are clearly documented and intentional.

The change is well-reasoned and the comments document the trade-offs clearly. The one gap is that all userStats UPDATE failures — including transient ones — are permanently dropped rather than retried, which could widen counter drift more than necessary.

apps/sim/lib/billing/core/usage-log.ts — specifically the catch block in recordUsage and any downstream callers that read currentPeriodCost / totalCost from userStats to enforce billing limits.

Important Files Changed

Sequence Diagram

sequenceDiagram
    participant Caller
    participant recordUsage
    participant DB_usageLog as DB: usage_log
    participant DB_userStats as DB: user_stats

    Caller->>recordUsage: recordUsage(params)

    alt "validEntries.length > 0"
        recordUsage->>DB_usageLog: INSERT rows (propagates errors)
        DB_usageLog-->>recordUsage: OK / throws
    end

    recordUsage->>DB_userStats: UPDATE userStats (lastActive, totalCost, currentPeriodCost)
    alt UPDATE succeeds, row found
        DB_userStats-->>recordUsage: "[{userId}]"
        recordUsage-->>Caller: void
    else UPDATE succeeds, row NOT found
        DB_userStats-->>recordUsage: []
        recordUsage->>recordUsage: logger.warn (counter dropped)
        recordUsage-->>Caller: void
    else UPDATE throws
        DB_userStats-->>recordUsage: error
        recordUsage->>recordUsage: logger.warn (counter dropped)
        recordUsage-->>Caller: void
    end

_{Reviews (1): Last reviewed commit: "fix(billing): drop transaction wrapper i..." | Re-trigger Greptile}