Confirmed live: /var/log/tinode.log-only redirection meant Loki never
received a single real tinode runtime log line (election, rehash,
connection, error output) since the container started -- Promtail
collects standard container stdout/stderr, not files written inside
the container. Every Pod's collected log was just this script's own
"NO_DB_INIT=true: skipping..." init line, nothing from tinode itself.
Now more consequential with native Cluster enabled: failover/rehash
state transitions were completely invisible to the normal Grafana/Loki
monitoring path, visible only via manual kubectl exec.
tee mirrors to both stdout (collected) and the file (still readable via
kubectl exec). PIPESTATUS[0] is required since a bare $? after a pipe
would report tee's exit status, not tinode's -- which also let this
drop the now-redundant "tail last 200 lines to stderr on crash" block,
since the full live stream already went to stdout throughout the run.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
StatefulSet's RollingUpdate replaces Pods one at a time in strict
sequence (unlike a Deployment, and unaffected by podManagementPolicy:
Parallel, which only governs initial scale-up/down) -- confirmed live:
run #152 restarted correctly (2/3 pods updated, then the 3rd) but hit
the old 120s ceiling before the last pod turned Ready, so CI reported
"failure" for a rollout that actually completed successfully seconds
later. 300s gives enough headroom for three sequential pod cycles
(image pull + start + native-Cluster rejoin) instead of tuning it to
the exact observed timing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
im-server moved from Deployment to StatefulSet in im-deploy-config
(native Cluster enablement) -- this workflow's kubectl commands still
targeted the old resource type, so every deploy since the migration
failed at the final step with "deployments.apps im-server not found"
(image built/pushed fine, just never got rolled out).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Complete the 4251-line docking document for IM contact us interface, including function description, interface definition, client rendering rules, local debugging example, operation configuration steps and reference implementation
Favorites from Drafty messages (link/file/image/video/voice) always fell back to text because imMessageType only recognized flat file_id/url content. Derive the type from Drafty ent[] entities: LN->link, IM->image, VD->video, AU->voice, EX refined by mime. Attachments win over inline links.
Logout left contacts seeing the user online until the me topic idle timeout
fired presUsersOfInterest(off), and even that frame was silently dropped
for p2p routes by handlePresence's xoriginal check. MarkUserOffline now
returns the remaining session count and cleanUp notifies users of interest
immediately once it hits zero; p2p presence is delivered straight to the
contact's me topic, mirroring presSingleUserOffline. Zentao #22.
The REST recall response reached only the caller, so peers kept showing the
recalled message until a full resync. broadcastIMMessageRecalled now pushes
the native {pres what:del delseq} tombstone through the hub for both single
and batch recall (empty pres topic for p2p, required by handlePresence's
xoriginal check); clients flush the message and render the recalled
placeholder. Zentao #16, #2.
Zentao #24/#25: commit 50f8581 added friend_request_verification and group_invite_verification to userPrivacyByUID, but existing databases lack those columns, so GetPrivacy failed with Error 1054 and every cross-user profile/qrcodes/friend-request endpoint returned 500.
Add user_privacy_schema.go which, under GET_LOCK, checks information_schema and ALTERs in any missing privacy columns, wired into the root adapter Open() and per-tenant TenantAdapterFor so tenant-scoped databases are provisioned too.
richjohnrambo