A checked-in snapshot attributes 59 of a Go service's 67 metric records to the platform

Add a queue to the manifest and three metric records appear with no instrumentation code in the service repository: amqp_delivery_total{queue,handler,status}, a delivery duration histogram, and an in-flight gauge ^[11]. A declared Postgres pool brings connection pool state from the driver, and a registered gRPC handler gets request counts, durations and statuses per method from the platform's interceptors ^[13][10]. All of it comes from the shared platform library the service imports, wired because the manifest declared the corresponding resource ^[15].

The counts run 59 records with source platform and 6 with source service ^[2][3]. On the stated total of 67 that is 88 percent platform and 9 percent service ^[21]. Brilliantov wrote that nine out of ten records exist "because the service was started, not because anyone decided to measure something" ^[5].

The published breakdown does not close. Fifty-nine platform records plus six service records plus the single <dynamic> record is 66, one short of the stated 67 ^[22][4]. The post does not place the remaining record.

A record is a metric name, not a time series. amqp_delivery_total carries three labels ^[11], so one line in that file becomes as many series at runtime as there are queue, handler and status combinations. The 59-to-6 split describes who owns the names. Storage and scrape cost track label cardinality, and a platform that ships a duration histogram per gRPC method is the bigger contributor there too ^[10].

The file is the part I would copy first. "You can't diff a scrape endpoint. You can diff a snapshot," Brilliantov wrote ^[17]. A metric that exists only in a running process cannot be reviewed in a pull request ^[23], so a new name becomes a line someone approves instead of something found later during an incident. He also wrote, "I'd rather start with a count, because the count is the argument" ^[16].

app_version is the entry that justifies itself on failure modes. It is one gauge carrying the version string linked into the binary ^[8]. The Go linker silently ignores -X pointed at a symbol that does not exist, so the build stays green, the version stays dev, and dev is what lands in the audit trail and the trace attributes ^[9]. Publishing it as a metric puts that on a dashboard. In the same spirit, the platform serves /health and /ready on the system port and the service implements neither handler, and there is no wait-for-health step in the deploy pipeline ^[6][7]. One fewer step to time out.

For the ratio to transfer, your services have to be shaped like this one: gRPC in, broker deliveries consumed, an outbound relay drained, Postgres behind it ^{[10][11][12][13]}. Six service metrics is what is left over when the platform owns transport, broker, pool and Go runtime ^[14]. A service with real domain counters would add more on its own side, and the split would move. You also need the library and the manifest first, and Brilliantov is describing a live PHP monolith being carved into Go services ^[18]. "This is what I do on one codebase, with the reasons and the price," he wrote ^[20].