[matthewdarwin]

Adds bstream_blockstream_subscriptions_closed_full_total, a counter of block stream subscriptions closed because they were full.

A subscriber (e.g. a relayer connected to a reader) that falls more blocks behind than the subscription channel size is disconnected (#91 made that size configurable). Until now this only showed as a reach max buffer size for subcription, closing channel log line, so there was nothing to graph or alert on.

Change

• blockstream/metrics.go: a bstream_blockstream metric set with one counter, SubscriptionsClosedFull.
• subscription.Push increments it inside the existing quitOnce, so each closed subscription counts once. Blocks pushed after the close aren't counted again.
• NewUnmanagedServer (and so NewServer) registers the set. A process exports the counter only when it creates a block stream server, and then from 0. dmetrics.Set.Register ignores repeat calls.
• No labels: the subscriber name comes from the client's request.
• The ServerOptionWithSubscriptionChannelSize CHANGELOG entry from blockstream: make subscription channel size configurable #91 appeared twice under Unreleased, so I removed the duplicate.

The cost is one increment when a subscription is dropped, nothing per block.

Tested

• TestBlockServerSubscriptionClosedWhenFull now also checks the counter: unchanged while the subscription holds up to its size, +1 when it closes, unchanged for blocks pushed after that.
• New TestBlockServerRegistersMetrics: creating a server exports the counter through prometheus.DefaultGatherer.
• End to end with firecore (v1.21.2 + this branch), reader-node-stdin + relayer with the default subscription size (200), a deliberately slow second subscriber and a 5,000-block burst: /metrics showed bstream_blockstream_subscriptions_closed_full_total 0 before the burst and 1 after, matching the single reach max buffer size log line.

🤖 Generated with Claude Code