nbari / pg_exporter / 32159669898

89%

main: 89%

Build Type

push

github

Committed by nbari

Commit Message

make graceful collector skips clear their metrics

A collector could finish a scrape successfully without publishing anything — an
unsupported server version, a missing view, a revoked privilege, an absent certificate
file — and its previous values stayed in the registry, served as though current. Only the
temp collector cleared them, and its history shows why a convention was not enough: temp
was copied from slru, inherited slru's missing reset, was fixed alone, and the fix never
propagated back.

## The trait

    fn collect_once(..) -> Result<Collected>   // implementation hook
    fn reset_metrics(&self)                    // required, no default
    fn collect(..) -> Result<()>               // provided: settles a Skipped

collect() keeps its name and Result<()> signature, so the registry, all 16 umbrellas and
all 283 integration-test call sites are untouched and every one of them now settles
automatically. Had settlement been a new method used only in production, the tests would
have exercised the unsettled path and could have passed while production behaved
differently. This is source-incompatible for anyone implementing Collector outside the
tree.

Skipped means "published nothing at all". A collector that refreshes part of its surface
and skips the rest must report Fresh, or the wrapper clears what it just published. An
error never clears, so a transient fault preserves registry state rather than destroying
it — which is why "I could not read my source" must not be reported as a skip.

## Scalars

A scalar metric cannot be removed while registered, so clearing one could only mean
zeroing, and a zeroed pg_ssl_certificate_valid ("invalid") or pg_ssl_enabled ("TLS off")
is a false alarm rather than an absence. Skip-capable scalars are therefore declared as
zero-label metric vectors: byte-identical on the wire, since prometheus emits no braces
for an empty label set, but removable via reset(). Pinned by a unit test.

## Bugs fixed along the way

... (continued)

Coverage Stats

1035 of 1490 new or added lines in 70 files covered. (69.46%)

348 existing lines in 28 files now uncovered.

18972 of 21150 relevant lines covered (89.7%)

426.68 hits per line