The window was carried across a spell, so the results either side of it ended up
adjacent: a monitor at [t, t, t, f, f] when egress dropped resumed at
[t, t, f, f, f] on its first real failure afterwards, crossing the threshold on
results taken before the outage.
Cleared in the producer beside the maintenance gate, which already clears the
window on entering a window, and guarded on its length so it is one write per
spell rather than one per degraded check. The monitor still goes down once the
rebuilt window holds enough real failures.
The two integration tests that asserted the old behaviour are rewritten: one
expected the window preserved through a spell, the other expected failures
either side of a spell to combine.
🔧
Distinguish "the peer answered" from "nothing answered", widen attribution to
every check that leaves the instance, make the maintenance exit explicit, and
fix the recovery poll cadence.
Providers can now say outright whether the peer answered, via peerResponded on
MonitorStatusResponse. Previously the producer inferred it from an HTTP status
code, which two providers do not report:
- GrpcProvider: NOT_SERVING shares NETWORK_ERROR with a connection failure, and
gRPC codes 0-16 are not HTTP statuses, so PERMISSION_DENIED, UNIMPLEMENTED
and the rest were all read as transport failures. Only DEADLINE_EXCEEDED and
UNAVAILABLE, or a failure carrying no gRPC code, now count as unreachable.
- DNSProvider: every resolver error became NETWORK_ERROR, so an authoritative
NXDOMAIN, SERVFAIL or REFUSED was read as unreachable. The c-ares codes where
the server replied are now separated from timeout and connection refused.
Left unset, the HTTP-status heuristic still applies, so other providers are
unchanged.
Attribution now follows whether the check leaves the instance, not the monitor
type. A hardware check is an ordinary outbound HTTP request to the Capture
agent, and a Docker daemon can be reached over TCP, so both are attributable;
only a Docker daemon on a unix socket is local and exempt.
Leaving a maintenance window is now explicit in the producer, alongside the
entry it already owned. It was a side effect of evaluating a check, which the
degraded-egress short-circuit skips, so a monitor whose window ended during an
outage kept reporting "maintenance" for the length of it.
The recovery poll interval is fixed at 5 seconds and no longer configurable.
While degraded, checks are flagged from stored state without probing, so the
interval is the window in which a recovered instance still discards its checks;
the only effect of raising it was a longer blind window. egressPollIntervalSeconds
is removed from the settings, the validator and the OpenAPI spec.
🔧
CheckProducer consults egress only for a transport failure: the check
failed, no HTTP status came back (isHttpStatusCode from PR 1), and the
monitor type is one whose failure can be a loss of egress
(EgressAttributableTypes; hardware and docker are excluded). A 4xx/5xx
or a content mismatch is evaluated as usual while the instance is
degraded. A rejected assessment is logged and treated as unknown so the
check is still buffered.
EXCLUDE_DEGRADED_EGRESS_MATCH moves out of the model into
check.query.ts, with an expression form for $group accumulators.
Daily status buckets and docker stats buckets exclude degraded checks
from their up/total counts while keeping them in the response-time
average, matching the uptime facet. toEntity maps egressStatus like its
siblings.
Adds heartbeatJobDegradedEgress integration tests, extends the
repository exclusion tests, and makes the harness egress stub settable.
🔧
CheckProducer asks EgressService to assess the instance's own connectivity
after a failing check and stamps the result on the check as egressStatus.
CheckEvaluator short-circuits degraded checks so no incident opens and no
spurious recovery fires later. Uptime and down-count aggregations exclude
degraded checks; the response-time series keeps them.
🔧
- The egress job runs through WorkerHelper.getEgressRecoveryJob(), like
the cleanup jobs, instead of a direct service call from the queue
- Toggling egressCheckEnabled resets the egress state and removes any
pending recovery job
- Probes resolve and pass the proxy, so http targets honour the global
proxy as monitors do
- PingProvider keeps IPv6 addresses intact instead of truncating at the
first colon
- Egress targets are parsed by one function shared by the validator and
the probe; malformed addresses such as unbalanced brackets are rejected
- Notification id pattern drops the /i flag, which produced an invalid
OpenAPI pattern
- EgressState findSingleton reads before it writes, so a read no longer
bumps updatedAt
- "disabled" is a distinct assessment result and is cached like any other
- Concurrent failures share an in-flight assessment for as long as it
runs, then reuse the settled result for 5 s
- The recovery job row is only removed if nobody rescheduled it since
this run claimed it, so a concurrent new episode keeps its job
- Deleting a notification removes it from egressNotifications
🔧
A failed check still calls assessAfterFailure, which now shares one
in-flight assessment between concurrent failures (same cache shape as
ProxyResolver). When egress degrades it upserts a global "egress" job;
DBQueueWorker runs checkRecovery for that type at the configured
interval, and the job deletes its own row once any target answers.
The unref'd timer, startRecoveryPolling, stop and init are gone: a
degraded episode survives a restart through the jobs collection.
deleteByIdAndType accepts a null refId for global rows.
🔧
Alex Holliday
Craig Lawson