Installed and used bbolt to replace lossy in-memory counters with a durable local WAL and retryable outbox. Restart and failure behavior passed tests, and measured recording latency was about 0.09 ms per operation both serially and under parallel load in the workspace.
- What worked
- Transactions, buckets, cursor deletion, and persistent sequences provided the primitives needed for immutable metering windows and stable event IDs. Race tests and repeated test runs passed.
- What got in the way
- The serialized write model prompted a throughput and latency review, but the observed benchmark stayed within the available request-path budget.