TRAM Roadmap

Planned features and known issues. Items are assigned to a version once scope is confirmed; unconfirmed work lives in the backlog at the bottom.

---

v1.2.3 — SNMP Poll v3 Validation & ASN.1 Decode Hardening

• SNMP poll source — validate SNMPv3 USM on real-device GET and WALK; keep existing walk / yield_rows coverage green
• ASN.1 serializer — decode-path coverage and wording updated for explicit decode-only behavior
• SNMP trap source — deferred; blocked by push-source architecture gap (issue #11, resolved in v1.3.0)
• SNMP trap sink — deferred until a reachable real receiver/test target is available

---

v1.3.0 — Broadcast Streams & Push-Source Scaling

Full design: docs/archive/v1.3.0-plan.md Skips v1.2.4–v1.2.7 due to severity of push-source architecture gap (issue #11). Scope is intentionally limited to HTTP push sources (webhook, prometheus_rw). UDP sources (syslog, snmp_trap) and dynamic K8s service provisioning are v1.3.1.

A — workers: block & multi-dispatch (issue #11)

• WorkersConfig Pydantic model — count: 1|all (v1.3.0 runtime) + count:N and list: forward-compatible in schema; validators: mutual exclusion, count >= 1, list non-empty + no duplicates; source-type defaults at validation (count: all for push HTTP, count: 1 for all others)
• BroadcastResult dataclass — accepted/rejected worker URLs, running | degraded | error status
• Linter rules L006–L010 — manager mode only; suppressed entirely in standalone; L007 fires on any multi-worker spec (count > 1, count: all, or list:) for poll/batch sources; no double-warning on queue sources (L009 for queue, L010 for non-queue); tram validate reads TRAM_MODE from env or --mode
• WorkerPool._worker_ids — worker_id → URL mapping from health poll
• WorkerPool.resolve() — v1.3.0: count:1 and count:all; count:N and list: raise NotImplementedError
• WorkerPool.multi_dispatch() — v1.3.0: count:1 and count:all paths
• WorkerPool.dispatch() — thin wrapper around multi_dispatch(count=1) for backwards compat
• Controller: _broadcast_placements, _active_placement_group, _make_placement_group_id(); degraded pipeline status value

B — Worker Public Ingress Security

• create_worker_ingress_app() — minimal FastAPI on :8767; webhooks only, no /agent/*
• Coupled shutdown: either listener thread exits → SIGTERM self → K8s restarts pod
• Composite GET /agent/health returns ok: false when ingress thread is dead
• TRAM_WORKER_INGRESS_PORT (default 8767); readiness probe stays on :8766 (composite)
• Worker StatefulSet: ingress containerPort 8767
• middleware.py import isolation contract enforced by test_worker_import_isolation.py

C — Unified Pipeline Stats & Load-Aware Dispatch

• PipelineStats dataclass (tram/agent/metrics.py) — thread-safe; records + bytes (bytes_in, bytes_out) + rolling error window; covers both stream and batch runs; replaces StreamMetrics
• ActiveRun.stats: PipelineStats for all run types
• executor.stream_run() and executor.batch_run() — stats param; bytes_in from source read, bytes_out from sink write; batch emits is_final: true stats report immediately before _post_run_complete
• RunCompletePayload gains bytes_in: int = 0 and bytes_out: int = 0; on_worker_run_complete writes them to run_history — this is the authoritative final total path
• Worker stats reporting thread (_stats_loop) — periodic reports is_final: false; batch completion emits is_final: true; posts every TRAM_STATS_INTERVAL (default 30s; replaces TRAM_HEARTBEAT_INTERVAL)
• POST /api/internal/pipeline-stats — is_final: true → StatsStore.remove(run_id) (load eviction only, not the bytes persistence path); otherwise StatsStore.update()
• StatsStore (tram/agent/stats_store.py) — keyed by run_id; staleness contract: get_by_run_id() stale-aware (reconciler + placement API per-slot); for_worker(), for_pipeline(), all_active() exclude stale (load-scoring and aggregate views); explicit remove(run_id); replaces HeartbeatStore
• on_worker_run_complete — writes bytes_in/bytes_out to run_history; calls StatsStore.remove(run_id) as fallback for crashed runs
• WorkerPool.load_score() — resolves worker_url → worker_id via _worker_ids before StatsStore.for_worker(); fallback to active_runs × 1 MB proxy

D — PlacementReconciler

• PlacementReconciler background thread (tram/agent/reconciler.py) — runs every min(TRAM_STATS_INTERVAL, 10)s; owns all stale detection, re-dispatch, and reconciling-window timeout; stats endpoint is write-only
• Stale slot detection — StatsStore.get_by_run_id(slot.current_run_id); age > 3 × interval → mark stale → re-dispatch same worker; updates slot.current_run_id in slots_json + DB (count:all only in v1.3.0)
• Reconciling-window timeout — after 2 × interval: matched → running; partial → degraded; none → re-dispatch
• Unit tests: stale → re-dispatch, reconciling timeout, partial recovery → degraded

E — Manager Restart Reconciliation

• DB table broadcast_placements — slots_json entries carry run_id_prefix (immutable) and current_run_id (mutable, updated on each re-dispatch)
• TramDB: save_broadcast_placement(), get_active_broadcast_placements(), update_broadcast_placement_status(), update_slot_run_id()
• _boot_load() seeds _broadcast_placements from DB; sets reconciling status
• Stats receiver matches incoming stats by run_id_prefix → resolves reconciling slot; updates current_run_id
• Unit tests: cold restart, partial recovery → degraded, unmatched → re-dispatch; current_run_id updated after re-dispatch

F — Placement API

• GET /api/pipelines/{name}/placement — iterates slots_json as source of truth; per-slot stats via StatsStore.get_by_run_id(slot.current_run_id); per-sec fields zeroed for stale; stale slots visible
• GET /api/cluster/streams — aggregate totals from StatsStore.for_pipeline() (non-stale) + group status/counts from broadcast_placements

G — Manager StatefulSet

• Replace manager-deployment.yaml with manager-statefulset.yaml + volumeClaimTemplates
• manager-headless-service.yaml for stable pod DNS
• Remove separate manager PVC from pvc.yaml; existingClaim migration path in values.yaml
• helm/NOTES.txt upgrade migration note

H — Alert cooldown on confirmed delivery only (issue #3)

• _fire_webhook() / _fire_email() return True/False
• _set_cooldown() called only on True
• Tests: HTTP 500, connection error, SMTP failure → no cooldown

---

v1.3.1 — Placement Completion & Targeted Backend Follow-ups

Depends on v1.3.0 TCP path being validated end-to-end.

• workers.count: N runtime — logical slot model (slot number ≠ worker assignment); PlacementReconciler spare-worker gap fill; slot reassignment on failover; WorkerPool.resolve() + multi_dispatch() N-worker paths
• workers.list: [...] runtime — named-worker placement; per-slot pinned re-dispatch on recovery
• Dynamic K8s Service provisioning — kubernetes: pipeline block; manager creates/deletes NodePort Services via K8s API; RBAC Role + RoleBinding; tram[k8s] optional extra
• source_stem / source_suffix filename tokens (issue #9) — added to all 6 file-based sinks
• Migrate SNMP to pysnmp 7.x (issue #10) — update import paths; pysmi migration

---

v1.3.2 — Metrics/Stats Parity & UDP Multi-Worker Streams

Follows v1.3.1 runtime stabilization. Scope is observability model cleanup plus UDP push-source validation.

• Standalone live stats parity — feed local active runs into StatsStore so standalone exposes the same live stream/placement-style stats model as manager mode instead of only history + process-local Prometheus metrics
• Manager operational metrics — add manager-side Prometheus series for dispatch/re-dispatch attempts, placement status counts, reconcile actions, worker health, and callback receipt (not direct failure counts — failures are inferred from the dispatch/receipt delta); document that /metrics is process-local and worker scraping is still required for cluster-wide execution metrics
• UDP multi-worker streams — syslog and snmp_trap sources; per-pipeline NodePort Service via kubernetes: enabled: true (shared selector for count: all, manual Endpoints for count: N / workers.list); kubernetes: enabled: true required in manager mode
• ASN.1 structured decode flattening — BER multi-record split for concatenated files; ordered message_classes fallback; generic json_flatten and hex_decode transforms with configurable ASN.1-oriented options and docs
• CDR record shaping — dotted-path support on unnest, explode, drop, rename, value_map, cast; shared path_utils helpers; select_from_list, coalesce_fields, and project primitives; conditional drop; and narrow single-segment wildcard support on hex_decode / json_flatten for explicit row-semantic CDR pipelines (LTE/SGW/PGW)
• Batch reconciliation — dedicated BatchReconciler, active batch lease tracking, worker-scan adoption after manager restart, and lost-run synthesis through the existing run-finalization path
• Incremental large-record batch processing — pipeline-level record_chunk_size, serializer parse_chunks(...) hook, and incremental ASN.1 BER decode for bounded-memory serial batch processing
• Safe staged batch file output — record-safe serial batch sinks (csv, ndjson) finalize output per source file, delete staged temp artifacts on failure, and support optional post_batch_cleanup heap trimming per pipeline

---

v1.3.3 — UI Revalidation & Backend Contract Sync

Follows v1.3.2 backend work. Scope is intentionally UI-heavy: revalidate every page against the shipped API and remove contract drift.

• Full UI/backend contract audit — verify every page, action, filter, export, and status badge against current API fields and route behavior
• Cluster and placement UX — finalize stream placement, slot health, stale/degraded/reconciling states, and manager-vs-standalone presentation
• Metrics/stats UX sync — align dashboard, cluster streams, run history, and Prometheus guidance with the actual manager/worker vs standalone stats model
• Pipeline detail completeness — schedule, alert, run-history, versioning, placement, and error-policy views all reflect current backend fields without fallback mismatches
• Plugin/templates/settings pages revalidation — remove stale assumptions, ensure live API-backed rendering, and verify empty/error states
• Responsive/browser pass — light/dark mode, mobile/tablet layout, console-clean build, and page-level smoke checks

---

Backlog (unversioned)

Connector Fixes (deferred from v1.2.4–v1.2.7)

• Kafka source — reconnect, offset commit, consumer group edge cases
• Kafka sink — producer error handling, retry, serializer integration
• OpenSearch sink — bulk write, index template, auth, retry on 429
• ClickHouse source/sink — query execution, batch insert, type coercion
• InfluxDB source/sink — line protocol, bucket/org resolution, token auth
• REST source/sink — auth types, pagination, retry, SSL verify
• gNMI source — subscription modes, path encoding, TLS
• SFTP source/sink — file glob, move-after-read, skip_processed, key auth
• FTP source/sink — passive mode, directory listing, file write
• S3 source/sink — bucket/prefix, multipart upload, credential chain
• MQTT source/sink — QoS levels, reconnect, topic wildcards
• AMQP source/sink — exchange/queue binding, ack/nack, prefetch
• NATS source/sink — subject routing, JetStream, reconnect

Operations & Observability

• Pipeline cloning — copy a pipeline with a name prompt in the UI
• Scheduled alert evaluation — cron-based alert checks independent of pipeline runs
• Dead-letter queue viewer — browse and replay DLQ records via the UI
• Per-sink record counts — run metrics broken down per sink
• Pipeline dependency graph — visualize pipeline chains when A feeds B
• Bulk actions — start/stop/delete multiple pipelines from the list view
• Live log streaming — WebSocket tail of log output for running stream pipelines
• Node health detail page — per-worker pipeline assignments and load in manager mode

Security & Multi-tenancy

• Role-based access — read-only vs admin token scopes (viewer/operator/admin)
• Per-pipeline API key scoping — restrict a key to specific pipelines
• Key upload API — POST /api/keys/upload / GET /api/keys / DELETE /api/keys/<name>
• Audit log — record who triggered, modified, or deleted pipelines

New Connectors & Serializers

• SMTP sink — outbound email delivery (alerts, reports)
• gRPC sink — generic gRPC unary call sink
• Syslog sink — forward records to remote syslog (RFC 5424)
• Kafka schema registry — full Avro + Protobuf with Confluent wire format
• PM-XML source — ingest 3GPP TS 32.435 PM XML files natively

Infrastructure

• Manager HA — standby manager with DB-backed leader election
• Graceful worker drain — POST /api/workers/{id}/drain; Helm pre-stop hook
• Coverage target increase — CI threshold raised to 75%; current coverage ~80%

---

Released

Version	Theme
v1.3.0	Broadcast streams; push-source scaling; placement reconciliation; manager StatefulSet
v1.2.3	SNMP Poll v3 validation; ASN.1 decode hardening
v1.2.2	Stability & polish; CI fixes; docs alignment
v1.2.1	Worker callback chain; run metrics propagation; dashboard UX
v1.2.0	Manager + Worker cluster architecture
v1.1.4	PipelineController; melt transform; SNMP fixes
v1.1.1	YAML diff modal; enrich missing file; template fixes
v1.1.0	Pipeline Wizard; Live Metrics; Alert Rules UI; AI Assist; Connector Test
v1.0.8/9	Browser auth; PostgreSQL subchart; shared RWX storage
v1.0.7	Bootstrap 5 web UI; dedicated K8s UI service
v1.0.6	ndjson serializer; per-sink serializer_out
v1.0.4	Schema registry; ClickHouse connector; pipeline update API
v1.0.3	MIB management API; schema management API; standard MIBs in image
v1.0.2	SNMPv3 USM (auth/priv)
v1.0.1	SNMP poll yield_rows; dynamic version
v1.0.0	API key auth; rate limiting; TLS; per-sink retry; circuit breaker; OTel
v0.9.0	Thread workers; batch_size; DLQ; CORBA source; processed-file tracking