ServiceRadar v1.4.65
The anomaly add-on no longer treats a cold start's first samples
as the level to measure drift against, drift findings report the
level the series actually moved to, and every Rust native add-on
now exits promptly on SIGTERM instead of outliving its RPC server.
Whats New
1.4.65
- anomaly add-on 0.3.11: the CUSUM drift anchor is captured only
once the rolling window holdsdrift_anchor_min_samplessamples
(default: the fullwindow_size, never belowmin_samples). An
add-on that had cold-started in the overnight trough anchored on
its first thirty samples and reported the ordinary morning ramp
as sustained upward drift on a flat uplink (PR #438). - anomaly add-on 0.3.11: drift verdicts carry
drift_level, the
mean raw value over the run in the metric's units. The former
reconstructiondrift_target + drift_shift_sigma * drift_scale
overstated the level several times over: the anchor scale is
refreshed every sample while its center stays frozen. The
add-on also writes its checkpoint on shutdown (PR #438). - Add-on SDK: after SIGINT or SIGTERM a Rust add-on runs its
shutdown hook, drains its gRPC server for at most two seconds,
and exits. Before, the drain waited for the agent's never-ending
telemetry stream, the process outlived its server, the supervisor
reported it unhealthy without restarting it, and every upgrade
ended in SIGKILL with no final checkpoint. netprobe 0.2.63,
otel-collector 0.1.7 and powerdns 0.1.8 carry the rebuilt SDK
(PR #438). - Web: the drift finding chart draws the reported
drift_levelas
the sustained level and keeps the sigma reconstruction only for
verdicts from add-ons older than 0.3.11 (PR #438). - Bumblebee: the catalog refresh worker is on by default
(bumblebeeCatalogRefresh.enabled, env
BUMBLEBEE_CATALOG_REFRESH_ENABLED), so a fresh deployment
materializes and promotes a0.1.0catalog at boot instead of
failing every scan on catalog load; the scan unit's state
directory is group-writable so the agent can write it
(bumblebee-scan add-on 0.1.6) (PR #440). - Bumblebee: a device posture now moves with its agent when the
agent is re-identified to another device, instead of staying on
the device it had when first seen (PR #441).
Upgrade Notes
- After a genuine first start, drift detection now waits for a full
rolling window (five hours at one sample a minute) before it can
open an episode. Restarts with a checkpoint do not reset that
clock. Setdrift_anchor_min_sampleslower on a profile to trade
that protection for earlier detection. - netprobe, otel-collector and powerdns are re-released only for
the SDK change; their behaviour is otherwise unchanged.