ServiceRadar v1.5.2

This release finishes hardening the durable edge pipeline. Agents now
recover their edge spool by themselves when a crash leaves it wedged.
Transient failures are retried, and records that never get written stay
visible: they are counted, can be inspected, and can be replayed. Core and
web-ng report the pipeline's status. On the StarRocks side, device pages
read a fraction of the data they used to, and core keeps each warehouse
node's memory configuration in line with what ServiceRadar needs. It also
adds the standalone NVR recording service, disabled by default.

Upgrade Notes

• StarRocks rebuilds timeseries_metrics once, on the first core
  start. The table is now sorted by metric, device and time, so each
  dashboard panel reads its own series instead of a whole day of every
  metric. StarRocks cannot change a sort key in place without rewriting
  the whole table at once, so core copies it into a new table one hour
  at a time, swaps the two and drops the old one. While it runs, the
  table takes about twice its normal object storage. Hours that fail for
  lack of memory are retried after a back-off. Loads continue throughout.
  The hourly rollup is rebuilt afterwards and catches up one day per
  refresh; until then rollup reads use the raw table. On the demo
  warehouse (about 55 days, 12M rows a day) the copy took about an hour
  (#6461).
• Core now sets StarRocks node memory itself. Every 2 minutes, one
  core replica sets datacache_mem_size (10%) and, on shared-data
  compute nodes, lake_metadata_cache_limit (sized from each node's
  process limit) through information_schema.be_configs. To do this the
  warehouse user needs UPDATE on be_configs and OPERATE or NODE
  on SYSTEM; root has both. Without them, core logs the refusal and
  keeps running. mem_limit and update_memory_limit_percent cannot be
  changed at runtime on StarRocks 4.1.7: set them in the node
  configuration (the k8s/starrocks profiles now use 70% and 30%) and
  restart. Core's health API reports each node's memory reserve and warns
  below 2.5 GiB (#6474, #6472, #6486).
• The k8s/starrocks shared-nothing profile lowers backend
  mem_limit from 90% to 70%. Re-applying it restarts the StarRocks
  pods (#6474).
• New migrations: 20261108190000 adds apply-failure fields to
  add-on statuses (#6446). 20261109170000, 20261109180000 and
  20261109190000 add the edge spool recovery operations, assignments
  and effects tables (#6469). 20261110032550 adds camera local
  publication intents (#6480).
• New JetStream stream: TELEMETRY_EDGE_RECORD_RECOVERY_V1
  (work-queue retention, no age expiry, one durable consumer) carries
  edge spool recovery. It is included in every sizing profile's file
  storage budget (#6469).
• Gateways need core to accept edge records. Before it publishes an
  agent's records, the gateway reads that agent's recovery coverage from
  core. While core is unreachable the gateway withholds the session with
  edge recovery coverage: recovery_coverage_unavailable, and agents
  keep the records in their spools and resend them (#6469).
• Edge-record stream policy is unchanged. TELEMETRY_EDGE_RECORD_V1_BULK
  and its audit stream stay on DiscardOld with a 24h MaxAge. A store
  refusal from NATS (max bytes exceeded) is now retried from the agent
  spool instead of being treated as a systemic failure (#6471, #6490).
• Argo CD installs: the migrations Job is now recreated on every
  sync, so migrations run on upgrade without manual intervention (#6456).
• Stale agent rows are now retired on schedule. Four cron workers
  never ran in the core_elx release: exposure reconciliation, stale
  agent pruning, remote-access recording reaping and version retention.
  They do now. Stale agent rows are marked unavailable, never deleted
  (#6455).
• The device page's event list covers the last 24 hours and matches
  device identifiers exactly. It used a wildcard match over the whole
  event table, which ran the warehouse out of memory (#6485).

Whats New

Durable edge pipeline

• Agents recover a wedged edge spool by themselves through a bounded
  recovery coordinator. A crash in the middle of an append no longer
  blocks the agent's only lane for good. Unsupported spool storage is
  left intact and reported (#6469).
• Agent health reports edge spool coverage gaps and wedged lanes (#6447).
• Core health reports the edge pipeline's status: records waiting to
  roll forward, the active, candidate and draining contract epochs, the
  components that do not yet hold a candidate, and whether the issuer
  key is present. Agents whose gateway lacks the active epoch are flagged
  (#6481).
• Settings > Cluster lists held dead letters by contract and reason, and
  can replay or purge a group (#6431).
• Transient edge-record failures are redelivered, and dead letters are
  counted by reason (#6417). Unleased sweep results are redelivered after
  a transient ingest failure (#6423).
• Agents finish edge-record sessions instead of cancelling them (#6420).
• Later contract registry epochs activate from the coordinator (#6418).
• Expired edge ledger and delivery-slot rows are pruned hourly (#6460).
• Device availability is written in one global lock order during sweep
  ingest and device merges, which removes a deadlock between the two
  (#6450, #6463).

StarRocks

• The device page's CPU, memory and disk panels read only their own
  series (#6461), and its events query no longer joins the alias catalog
  or matches with LIKE (#6485).
• Settings > Cluster shows each StarRocks node's memory reserve, low-reserve
  warnings and keys that need a restart (#6486).
• The SRQL parity harnesses apply the warehouse schema through the
  migrator, so they wait for optimize jobs like a real upgrade does (#6479).
• Flow details read process attribution from StarRocks (#6408).

Recording and cameras

• New standalone serviceradar-nvr recording release and StatefulSet,
  disabled by default (nvr.enabled: false). It rotates bounded
  recording blocks, admits workers against spool budget and free space,
  recovers and seals interrupted blocks after a restart, serves verified
  media reads, and publishes sealed blocks to core through immutable
  publication intents (#6449, #6452, #6457, #6459, #6462, #6467, #6473,
  #6475, #6477, #6480, #6482, #6484, #6491).
• Camera relay stop reconciliation API and reconciliation of historical
  relay open outcomes (#6411, #6419). Relay open retries are deduplicated
  (#6414).

Other fixes

• helm install --wait no longer hangs, and CNPG secrets carry Helm
  ownership (#6407).
• Armis syncs are bounded by timeouts, and aborted runs are reported as
  failed syncs (#6405). Sync failures keep the full upstream error (#6410).
• Plugin and add-on downloads retry after timeouts. Add-on install
  failures are surfaced, and paused native add-on rollouts recover
  (#6415, #6446).
• Agent config section failures are scoped and reported per section
  (#6451).
• Sync status takes the IntegrationSource id only from sync metadata
  (#6448).
• Topology projects CONTAINS edges from virtualization inventory (#6454).
• The MTR daily comparison stays aligned just after midnight UTC (#6476).
• Agent release imports and publishing run in the background (#6409).