@obsunified/collector 2.0.1

package/src/migrations/009_projects_and_keys.sql

@@ -0,0 +1,22 @@
+CREATE TABLE IF NOT EXISTS projects (
+  id TEXT PRIMARY KEY,
+  name TEXT NOT NULL,
+  slug TEXT NOT NULL UNIQUE,
+  created_at TEXT NOT NULL
+);
+
+CREATE INDEX IF NOT EXISTS idx_projects_slug ON projects (slug);
+
+CREATE TABLE IF NOT EXISTS ingest_keys (
+  id TEXT PRIMARY KEY,
+  project_id TEXT NOT NULL,
+  key_hash TEXT NOT NULL UNIQUE,
+  key_prefix TEXT NOT NULL,
+  name TEXT NOT NULL,
+  created_at TEXT NOT NULL,
+  revoked_at TEXT,
+  FOREIGN KEY (project_id) REFERENCES projects(id)
+);
+
+CREATE INDEX IF NOT EXISTS idx_ingest_keys_project ON ingest_keys (project_id, revoked_at);
+CREATE INDEX IF NOT EXISTS idx_ingest_keys_hash ON ingest_keys (key_hash);

package/src/migrations/010_signal_project_id.sql

@@ -0,0 +1,15 @@
+-- Add project_id to every signal table. DEFAULT 'default' backfills existing rows.
+ALTER TABLE telemetry_spans ADD COLUMN project_id TEXT NOT NULL DEFAULT 'default';
+ALTER TABLE logs ADD COLUMN project_id TEXT NOT NULL DEFAULT 'default';
+ALTER TABLE usage_events ADD COLUMN project_id TEXT NOT NULL DEFAULT 'default';
+ALTER TABLE ai_calls ADD COLUMN project_id TEXT NOT NULL DEFAULT 'default';
+ALTER TABLE session_replay_metadata ADD COLUMN project_id TEXT NOT NULL DEFAULT 'default';
+ALTER TABLE user_profiles ADD COLUMN project_id TEXT NOT NULL DEFAULT 'default';
+
+-- Composite indexes for project-scoped queries.
+CREATE INDEX IF NOT EXISTS idx_spans_project_received ON telemetry_spans (project_id, received_at DESC);
+CREATE INDEX IF NOT EXISTS idx_logs_project_received ON logs (project_id, received_at DESC);
+CREATE INDEX IF NOT EXISTS idx_usage_project_received ON usage_events (project_id, received_at DESC);
+CREATE INDEX IF NOT EXISTS idx_ai_project_received ON ai_calls (project_id, received_at DESC);
+CREATE INDEX IF NOT EXISTS idx_replay_project_last ON session_replay_metadata (project_id, last_chunk_at DESC);
+CREATE INDEX IF NOT EXISTS idx_users_project_last ON user_profiles (project_id, last_seen_at DESC);

package/src/migrations/011_default_project_seed.sql

@@ -0,0 +1,3 @@
+-- Seed the default project. Existing rows already carry project_id='default' via migration 010's column default.
+INSERT OR IGNORE INTO projects (id, name, slug, created_at)
+VALUES ('default', 'Default', 'default', datetime('now'));

package/src/migrations/012_alert_rules.sql

@@ -0,0 +1,17 @@
+CREATE TABLE IF NOT EXISTS alert_rules (
+  id TEXT PRIMARY KEY,
+  project_id TEXT NOT NULL,
+  name TEXT NOT NULL,
+  signal TEXT NOT NULL,
+  query_json TEXT NOT NULL,
+  threshold REAL NOT NULL,
+  window_mins INTEGER NOT NULL,
+  comparison TEXT NOT NULL,
+  channels_json TEXT NOT NULL,
+  enabled INTEGER NOT NULL DEFAULT 1,
+  created_at TEXT NOT NULL,
+  updated_at TEXT NOT NULL,
+  FOREIGN KEY (project_id) REFERENCES projects(id)
+);
+
+CREATE INDEX IF NOT EXISTS idx_alert_rules_project ON alert_rules (project_id, enabled);

package/src/migrations/013_alert_evaluations.sql

@@ -0,0 +1,13 @@
+CREATE TABLE IF NOT EXISTS alert_evaluations (
+  id TEXT PRIMARY KEY,
+  rule_id TEXT NOT NULL,
+  project_id TEXT NOT NULL,
+  evaluated_at TEXT NOT NULL,
+  value REAL NOT NULL,
+  state TEXT NOT NULL,
+  notified INTEGER NOT NULL DEFAULT 0,
+  FOREIGN KEY (rule_id) REFERENCES alert_rules(id)
+);
+
+CREATE INDEX IF NOT EXISTS idx_alert_eval_rule_time ON alert_evaluations (rule_id, evaluated_at DESC);
+CREATE INDEX IF NOT EXISTS idx_alert_eval_project ON alert_evaluations (project_id, evaluated_at DESC);

package/src/migrations/014_alert_state.sql

@@ -0,0 +1,9 @@
+CREATE TABLE IF NOT EXISTS alert_state (
+  rule_id TEXT PRIMARY KEY,
+  project_id TEXT NOT NULL,
+  current_state TEXT NOT NULL,
+  last_state_change TEXT NOT NULL,
+  FOREIGN KEY (rule_id) REFERENCES alert_rules(id)
+);
+
+CREATE INDEX IF NOT EXISTS idx_alert_state_project ON alert_state (project_id);

package/src/migrations/015_metrics.sql

@@ -0,0 +1,58 @@
+-- OTLP metrics storage.
+--
+-- Two tables:
+--   metric_series — one row per unique (resource, scope, metric name, metric
+--                   attrs) tuple. Upserted on ingest.
+--   metric_point  — one row per datapoint. FK to metric_series.
+--
+-- Histogram-specific fields on metric_point are nullable (only populated for
+-- type='histogram'). Gauge/sum write `value`. Exponential histograms and
+-- Summary are not yet supported (Phase 5).
+
+CREATE TABLE IF NOT EXISTS metric_series (
+  id TEXT PRIMARY KEY,
+  project_id TEXT NOT NULL,
+  name TEXT NOT NULL,
+  description TEXT,
+  unit TEXT,
+  type TEXT NOT NULL,                 -- 'gauge' | 'sum' | 'histogram'
+  is_monotonic INTEGER,               -- 0/1 — sum only
+  temporality INTEGER,                -- 1=DELTA, 2=CUMULATIVE — sum/histogram
+  scope_name TEXT,
+  scope_version TEXT,
+  service_name TEXT,
+  resource_attrs_json TEXT,
+  attributes_json TEXT,
+  identity TEXT NOT NULL,             -- canonical JSON of identity tuple
+  created_at TEXT NOT NULL,
+  UNIQUE(project_id, identity)
+);
+
+CREATE INDEX IF NOT EXISTS idx_metric_series_project_name
+  ON metric_series (project_id, name);
+
+CREATE TABLE IF NOT EXISTS metric_point (
+  id TEXT PRIMARY KEY,
+  series_id TEXT NOT NULL,
+  project_id TEXT NOT NULL,           -- denormalized for retention sweep
+  ts_ns TEXT NOT NULL,                -- uint64 as decimal string
+  start_ts_ns TEXT,
+  value REAL,                         -- gauge/sum scalar
+  count INTEGER,                      -- histogram
+  sum REAL,                           -- histogram
+  min REAL,                           -- histogram
+  max REAL,                           -- histogram
+  bounds_json TEXT,                   -- histogram bucket upper bounds (number[])
+  bucket_counts_json TEXT,            -- histogram bucket counts (number[])
+  exemplars_json TEXT,                -- array of {value, traceId?, spanId?, tsNs}
+  received_at TEXT NOT NULL,
+  expires_at TEXT NOT NULL,
+  FOREIGN KEY (series_id) REFERENCES metric_series(id)
+);
+
+CREATE INDEX IF NOT EXISTS idx_metric_point_series_ts
+  ON metric_point (series_id, ts_ns);
+CREATE INDEX IF NOT EXISTS idx_metric_point_expires
+  ON metric_point (expires_at);
+CREATE INDEX IF NOT EXISTS idx_metric_point_project_received
+  ON metric_point (project_id, received_at DESC);

package/src/migrations/016_metrics_exp_histogram_summary.sql

@@ -0,0 +1,5 @@
+-- Extend metric_point with a JSON blob for type-specific fields that don't
+-- map cleanly onto columns: exponential histogram buckets (scale,
+-- zeroCount, positive/negative) and summary quantileValues. Keeps the
+-- schema compact without migrating whenever OTel adds a new metric type.
+ALTER TABLE metric_point ADD COLUMN extra_json TEXT;

package/src/migrations/017_span_dropped_counts.sql

@@ -0,0 +1,11 @@
+-- OTLP span-level passthrough columns.
+--
+-- Matches RFC 0001 schema-conformance bar: droppedAttributesCount,
+-- droppedEventsCount, droppedLinksCount, and W3C traceState must survive the
+-- receiver end-to-end. Event/link-level dropped counts live inside
+-- events_json / links_json already.
+
+ALTER TABLE telemetry_spans ADD COLUMN trace_state TEXT;
+ALTER TABLE telemetry_spans ADD COLUMN dropped_attributes_count INTEGER NOT NULL DEFAULT 0;
+ALTER TABLE telemetry_spans ADD COLUMN dropped_events_count INTEGER NOT NULL DEFAULT 0;
+ALTER TABLE telemetry_spans ADD COLUMN dropped_links_count INTEGER NOT NULL DEFAULT 0;

package/src/migrations/018_log_flags.sql

@@ -0,0 +1,8 @@
+-- OTLP log-record passthrough columns.
+--
+-- Matches RFC 0001 schema-conformance bar: LogRecord.flags (W3C trace
+-- flags bitfield) and droppedAttributesCount must survive the receiver
+-- end-to-end.
+
+ALTER TABLE logs ADD COLUMN flags INTEGER NOT NULL DEFAULT 0;
+ALTER TABLE logs ADD COLUMN dropped_attributes_count INTEGER NOT NULL DEFAULT 0;

package/src/migrations/019_ai_span_payloads.sql

@@ -0,0 +1,30 @@
+-- Side table for large prompt/completion/tool/retrieval payloads attached
+-- to OpenInference-kind spans (LLM/CHAIN/RETRIEVER/EMBEDDING/TOOL/AGENT/...).
+--
+-- Spans themselves continue to live in telemetry_spans — this table only
+-- holds the input/output blobs that would otherwise bloat the hot spans
+-- table. Keyed by (trace_id, span_id) so a join against telemetry_spans
+-- reconstructs the full AI span.
+--
+-- Retention mirrors telemetry_spans: entries are purged when expires_at
+-- elapses. span_kind is denormalized from attributes_json for cheap
+-- filtering without json_extract.
+
+CREATE TABLE IF NOT EXISTS ai_span_payloads (
+  project_id TEXT NOT NULL,
+  trace_id TEXT NOT NULL,
+  span_id TEXT NOT NULL,
+  span_kind TEXT NOT NULL,
+  input_json TEXT,
+  output_json TEXT,
+  received_at TEXT NOT NULL,
+  expires_at TEXT NOT NULL,
+  PRIMARY KEY (trace_id, span_id)
+);
+
+CREATE INDEX IF NOT EXISTS idx_ai_payloads_received
+  ON ai_span_payloads (received_at DESC);
+CREATE INDEX IF NOT EXISTS idx_ai_payloads_project
+  ON ai_span_payloads (project_id, received_at DESC);
+CREATE INDEX IF NOT EXISTS idx_ai_payloads_kind
+  ON ai_span_payloads (project_id, span_kind, received_at DESC);

package/src/migrations/020_ai_span_evaluations.sql

@@ -0,0 +1,33 @@
+-- Evaluations attached to an AI span. One span can have many evaluations
+-- (e.g. an LLM-judge "hallucination" score plus a human "thumbs_up" label).
+--
+-- source tells us where the evaluation came from:
+--   "llm_judge"  — another model scored this span
+--   "code"       — deterministic check (regex, schema, etc.)
+--   "human"      — UI annotation / feedback
+--   "user"       — end-user thumbs up/down in the product
+--
+-- score is numeric (0..1 or any float); label is a free-form categorical
+-- (e.g. "correct", "hallucinated", "👍"). Either or both may be set.
+
+CREATE TABLE IF NOT EXISTS ai_span_evaluations (
+  evaluation_id TEXT PRIMARY KEY,
+  project_id TEXT NOT NULL,
+  trace_id TEXT NOT NULL,
+  span_id TEXT NOT NULL,
+  name TEXT NOT NULL,
+  score REAL,
+  label TEXT,
+  explanation TEXT,
+  source TEXT NOT NULL,
+  metadata_json TEXT,
+  created_at TEXT NOT NULL,
+  expires_at TEXT NOT NULL
+);
+
+CREATE INDEX IF NOT EXISTS idx_ai_evals_span
+  ON ai_span_evaluations (trace_id, span_id);
+CREATE INDEX IF NOT EXISTS idx_ai_evals_project_created
+  ON ai_span_evaluations (project_id, created_at DESC);
+CREATE INDEX IF NOT EXISTS idx_ai_evals_project_name
+  ON ai_span_evaluations (project_id, name, created_at DESC);

package/src/migrations/021_ai_span_session.sql

@@ -0,0 +1,13 @@
+-- Denormalize session.id and user.id off the span's attributes_json into
+-- ai_span_payloads so we can filter sessions / user threads cheaply without
+-- a json_extract on every row.
+--
+-- Both columns are nullable — not every AI span belongs to a session.
+
+ALTER TABLE ai_span_payloads ADD COLUMN session_id TEXT;
+ALTER TABLE ai_span_payloads ADD COLUMN user_id TEXT;
+
+CREATE INDEX IF NOT EXISTS idx_ai_payloads_session
+  ON ai_span_payloads (project_id, session_id, received_at DESC);
+CREATE INDEX IF NOT EXISTS idx_ai_payloads_user
+  ON ai_span_payloads (project_id, user_id, received_at DESC);

package/src/migrations/022_span_log_session_id.sql

@@ -0,0 +1,13 @@
+-- Denormalize session.id off the span/log attributes so we can pivot from a
+-- session to every telemetry signal (spans, logs) without a json_extract on
+-- every row. Populated at ingest by the enrichment plugins.
+--
+-- Nullable — not every span or log belongs to a session.
+
+ALTER TABLE telemetry_spans ADD COLUMN session_id TEXT;
+ALTER TABLE logs ADD COLUMN session_id TEXT;
+
+CREATE INDEX IF NOT EXISTS idx_spans_session
+  ON telemetry_spans (project_id, session_id, received_at DESC);
+CREATE INDEX IF NOT EXISTS idx_logs_session
+  ON logs (project_id, session_id, received_at DESC);

package/src/migrations/023_service_map_indexes.sql

@@ -0,0 +1,32 @@
+-- Service-map performance indices.
+--
+-- Two queries blow up at ~100k+ spans without these:
+--   1. The whole service-map endpoint filters by project_id + received_at
+--      window, but the existing idx_spans_received_at is single-column
+--      (received_at DESC) and can't prune on project_id, so multi-tenant
+--      installs scan the world.
+--   2. The new async-edge query (added in 522424e) gates on
+--      `links_json IS NOT NULL AND links_json != '[]'`. Without an index
+--      that knows about that predicate, every row gets a json_each call
+--      even when the column is NULL — order-of-magnitude slower than it
+--      needs to be.
+--
+-- Both indices target (project_id, received_at DESC) as the leading
+-- columns so range queries hit the index first and project filtering is
+-- free. The partial index for links is much smaller than the full one
+-- (links_json is non-empty on a small fraction of spans) so the planner
+-- prefers it for the link-follow path.
+
+CREATE INDEX IF NOT EXISTS idx_spans_project_received
+	ON telemetry_spans (project_id, received_at DESC);
+
+CREATE INDEX IF NOT EXISTS idx_spans_with_links
+	ON telemetry_spans (project_id, received_at DESC)
+	WHERE links_json IS NOT NULL AND links_json != '[]';
+
+-- Same scoping for service-name lookups: the per-service operations
+-- query (used by the click-through drawer) filters by
+-- (project_id, service_name, received_at). Existing idx_spans_service
+-- is service_name-only.
+CREATE INDEX IF NOT EXISTS idx_spans_project_service_received
+	ON telemetry_spans (project_id, service_name, received_at DESC);

package/src/migrations/024_analysis_results.sql

@@ -0,0 +1,49 @@
+-- RFC 0002 Stage 1: persistent storage for Analysis runs.
+--
+-- One row per (analysis_id, generated_at). The dashboard reads "latest result
+-- per analysis" via the (project_id, analysis_id, generated_at DESC) index;
+-- the retention cron drops rows past expires_at. Stage 1 only writes
+-- status / primary / baseline / payload; narrative + narrative_signature
+-- get filled by Stage 3.
+
+CREATE TABLE IF NOT EXISTS analysis_definitions (
+	project_id TEXT NOT NULL,
+	id TEXT NOT NULL,
+	title TEXT NOT NULL,
+	"group" TEXT NOT NULL,
+	source TEXT NOT NULL,                  -- 'tier0' | 'tier1' | 'user' | 'llm-suggested'
+	view TEXT NOT NULL,                    -- 'tile' | 'page' | 'alert'
+	refresh_seconds INTEGER,
+	sql TEXT,
+	scope_json TEXT,                       -- JSON-serialized scope object
+	created_at TEXT NOT NULL,
+	updated_at TEXT NOT NULL,
+	last_run_at TEXT,
+	PRIMARY KEY (project_id, id)
+);
+
+CREATE INDEX IF NOT EXISTS idx_analysis_definitions_group
+	ON analysis_definitions (project_id, "group", id);
+
+CREATE TABLE IF NOT EXISTS analysis_results (
+	id INTEGER PRIMARY KEY AUTOINCREMENT,
+	project_id TEXT NOT NULL,
+	analysis_id TEXT NOT NULL,
+	generated_at INTEGER NOT NULL,         -- unix epoch millis
+	params_hash TEXT,
+	status TEXT NOT NULL,                  -- 'ok' | 'warn' | 'critical' | 'unknown'
+	primary_value REAL,
+	baseline_value REAL,
+	delta_pct REAL,
+	payload_json TEXT NOT NULL,
+	narrative TEXT,
+	narrative_signature TEXT,
+	duration_ms INTEGER NOT NULL,
+	expires_at INTEGER NOT NULL            -- unix epoch millis; matched by retention cron
+);
+
+CREATE INDEX IF NOT EXISTS idx_analysis_results_latest
+	ON analysis_results (project_id, analysis_id, generated_at DESC);
+
+CREATE INDEX IF NOT EXISTS idx_analysis_results_expires
+	ON analysis_results (expires_at);

package/src/migrations/025_stage6_autopinning_alerts.sql

@@ -0,0 +1,38 @@
+-- RFC 0002 Stage 6: auto-pinning + alert→analysis binding.
+--
+-- Two coupled changes that share a migration so first-time installs land
+-- in one shot:
+--
+-- 1. ask_evidence_events
+--    Every time the Ask box's tool-use loop cites an analysis in its
+--    answer, we drop a row here. The Health tab's "Pinned" group derives
+--    from the top-cited analyses over the trailing 7 days. This is the
+--    auto-pinning input — pure observation, no manual config required.
+--
+-- 2. alert_rules.analysis_id
+--    Lets an alert rule reference an Analysis by id rather than carrying
+--    its own query. The evaluator reads the analysis's latest primary
+--    value; webhook payloads attach the analysis's current narrative so
+--    Slack/PagerDuty messages are about *what's happening*, not just
+--    "metric crossed threshold." NULL means the rule still uses its
+--    `query_json` (the pre-Stage-6 path).
+
+CREATE TABLE IF NOT EXISTS ask_evidence_events (
+	id INTEGER PRIMARY KEY AUTOINCREMENT,
+	project_id TEXT NOT NULL,
+	analysis_id TEXT NOT NULL,
+	asked_at INTEGER NOT NULL              -- unix epoch millis
+);
+
+CREATE INDEX IF NOT EXISTS idx_ask_evidence_events_recent
+	ON ask_evidence_events (project_id, asked_at DESC);
+
+CREATE INDEX IF NOT EXISTS idx_ask_evidence_events_analysis
+	ON ask_evidence_events (project_id, analysis_id, asked_at DESC);
+
+-- ALTER TABLE ... ADD COLUMN is the single supported schema-change form
+-- in SQLite/D1. Defaults to NULL (legacy rules unaffected).
+ALTER TABLE alert_rules ADD COLUMN analysis_id TEXT;
+
+CREATE INDEX IF NOT EXISTS idx_alert_rules_analysis
+	ON alert_rules (project_id, analysis_id);

package/src/migrations/026_obs_dashboard_project_seed.sql

@@ -0,0 +1,18 @@
+-- Seed a separate `obs-dashboard` project for the dashboard's own self-telemetry,
+-- so dogfooded page_views / errors / interactions don't pollute the `default`
+-- project that holds real demo + customer traffic.
+INSERT OR IGNORE INTO projects (id, name, slug, created_at)
+VALUES ('obs-dashboard', 'Obs Dashboard', 'obs-dashboard', datetime('now'));
+
+-- Deterministic ingest key whose plaintext lives in apps/web/.env.development.local
+-- (VITE_OBS_INGEST_KEY). Re-running this migration is a no-op via INSERT OR IGNORE
+-- on the unique key_hash.
+INSERT OR IGNORE INTO ingest_keys (id, project_id, key_hash, key_prefix, name, created_at)
+VALUES (
+  'obs-dashboard-bootstrap',
+  'obs-dashboard',
+  '6aa18dc1261b312e98b236d5a7dae698d1a90ecbe33d8024744063a066a35b9b',
+  'obs_dashboard_65',
+  'bootstrap',
+  datetime('now')
+);

package/src/migrations/027_identity_propagation.sql

@@ -0,0 +1,50 @@
+-- RFC 0004 — identity propagation.
+--
+-- Adds the `interaction_id` correlation key (a ULID minted by
+-- @obs/analytics-sdk at click/submit/keydown time) to every signal that
+-- carries it, plus the missing `session_id` column on `ai_calls`.
+--
+-- Why interaction_id: today, replay → trace navigation is approximate
+-- (session_id + timestamp window). With a click-scoped correlation key we
+-- can pivot from any rrweb event to the exact trace it caused, in either
+-- direction.
+--
+-- metric_point intentionally does NOT get either column. Metrics aggregate;
+-- tying a metric point to one click defeats the purpose. Exemplars
+-- (already supported via metric_point.exemplars_json) are the correct
+-- correlation primitive for metrics. Indexing exemplars for reverse lookup
+-- is a follow-up flagged in the RFC.
+
+ALTER TABLE telemetry_spans   ADD COLUMN interaction_id TEXT;
+ALTER TABLE logs              ADD COLUMN interaction_id TEXT;
+ALTER TABLE usage_events      ADD COLUMN interaction_id TEXT;
+ALTER TABLE ai_calls          ADD COLUMN interaction_id TEXT;
+ALTER TABLE ai_span_payloads  ADD COLUMN interaction_id TEXT;
+
+-- Backfill the one signal table missing session_id. ai_calls today
+-- carries trace_id but no session, forcing a two-hop join through spans
+-- to recover the user session that triggered an LLM call.
+ALTER TABLE ai_calls          ADD COLUMN session_id TEXT;
+
+-- Partial indices keep the index small in the common case where most
+-- rows predate this migration and therefore have NULL interaction_id.
+-- SQLite/D1 supports partial indexes via the WHERE clause.
+CREATE INDEX IF NOT EXISTS idx_spans_interaction
+  ON telemetry_spans (project_id, interaction_id, received_at DESC)
+  WHERE interaction_id IS NOT NULL;
+
+CREATE INDEX IF NOT EXISTS idx_logs_interaction
+  ON logs (project_id, interaction_id, received_at DESC)
+  WHERE interaction_id IS NOT NULL;
+
+CREATE INDEX IF NOT EXISTS idx_usage_interaction
+  ON usage_events (project_id, interaction_id, received_at DESC)
+  WHERE interaction_id IS NOT NULL;
+
+CREATE INDEX IF NOT EXISTS idx_ai_calls_interaction
+  ON ai_calls (project_id, interaction_id, received_at DESC)
+  WHERE interaction_id IS NOT NULL;
+
+CREATE INDEX IF NOT EXISTS idx_ai_calls_session
+  ON ai_calls (project_id, session_id, received_at DESC)
+  WHERE session_id IS NOT NULL;

package/src/migrations/028_profile_blobs.sql

@@ -0,0 +1,57 @@
+-- RFC 0007 — pprof profiling receiver storage.
+--
+-- Two tables:
+--   profile_blobs        — one row per ingested pprof blob (cpu / heap /
+--                          off-cpu / mutex / etc). The blob itself lives
+--                          in R2 (Workers) or filesystem (Node), keyed by
+--                          blob_url; this table is the metadata index.
+--   profile_trace_index  — one row per (profile, trace) pair so the
+--                          trace waterfall can join in O(1) on trace_id.
+--                          Populated at ingest from the
+--                          x-obs-trace-ids header (Phase 4 minimal) or
+--                          by parsing the pprof on later iterations.
+--
+-- The blob index uses an explicit join table rather than a JSON column
+-- so a 60s profile from a high-traffic service touching thousands of
+-- traces doesn't blow past D1's 2 MB row limit (see RFC 0007 §
+-- profile_trace_index rationale).
+
+CREATE TABLE IF NOT EXISTS profile_blobs (
+  id TEXT PRIMARY KEY,                    -- ULID minted at ingest
+  project_id TEXT NOT NULL,
+  service_name TEXT,
+  profile_type TEXT NOT NULL,             -- 'cpu' | 'heap' | 'wall' | 'block' | 'mutex' | 'goroutine' | 'offcpu'
+  start_ts TEXT NOT NULL,                 -- ISO8601 — beginning of sampled window
+  end_ts TEXT NOT NULL,
+  duration_ms INTEGER NOT NULL,
+  blob_size_bytes INTEGER NOT NULL,
+  blob_url TEXT NOT NULL,                 -- R2 key or filesystem path
+  sample_count INTEGER,                   -- total samples (nullable — header-driven path doesn't know)
+  agent TEXT,                             -- 'datadog-pprof' | 'pyroscope' | 'parca-agent' | 'otel-ebpf' | unknown
+  resource_attrs_json TEXT,               -- service.name, host.name, etc.
+  received_at TEXT NOT NULL,
+  expires_at TEXT NOT NULL
+);
+
+CREATE INDEX IF NOT EXISTS idx_profile_blobs_service_ts
+  ON profile_blobs (project_id, service_name, end_ts DESC);
+
+CREATE INDEX IF NOT EXISTS idx_profile_blobs_expires
+  ON profile_blobs (expires_at);
+
+-- Trace→profile join. One row per (profile, trace) pair. ~50 bytes per row;
+-- a profile sampling 5,000 distinct traces produces ~250 KB of join-table
+-- rows, indexed for sub-ms lookup.
+CREATE TABLE IF NOT EXISTS profile_trace_index (
+  profile_id TEXT NOT NULL,
+  trace_id TEXT NOT NULL,
+  project_id TEXT NOT NULL,               -- denormalized for retention sweep
+  PRIMARY KEY (profile_id, trace_id),
+  FOREIGN KEY (profile_id) REFERENCES profile_blobs(id) ON DELETE CASCADE
+);
+
+CREATE INDEX IF NOT EXISTS idx_profile_trace_index_trace
+  ON profile_trace_index (project_id, trace_id);
+
+CREATE INDEX IF NOT EXISTS idx_profile_trace_index_profile
+  ON profile_trace_index (profile_id);

package/src/migrations/029_telemetry_sdk_name.sql

@@ -0,0 +1,20 @@
+-- RFC 0009 Phase 5.3 — denormalize telemetry.sdk.name onto telemetry_spans.
+--
+-- The service map dashboard wants a "show me only Beyla-derived edges"
+-- filter. Doing that with json_extract on resource_attributes_json on
+-- every aggregation row is too slow (no index possible on JSON paths
+-- in D1). Denormalize at ingest into a top-level column with a partial
+-- index — partial because most spans will continue to come from
+-- regular SDKs and don't need the index entry.
+--
+-- Populated by default-span-enrichment from
+-- resource_attributes['telemetry.sdk.name']. NULL when the producer
+-- didn't set it (older SDKs); the service-map filter treats NULL as
+-- "sdk-derived" (the conservative default) so existing edges keep
+-- rendering when "sdk" is selected.
+
+ALTER TABLE telemetry_spans ADD COLUMN telemetry_sdk_name TEXT;
+
+CREATE INDEX IF NOT EXISTS idx_spans_sdk_name
+  ON telemetry_spans (project_id, telemetry_sdk_name, received_at DESC)
+  WHERE telemetry_sdk_name IS NOT NULL;

package/src/migrations/030_analysis_claim.sql

@@ -0,0 +1,18 @@
+-- RFC 0002 Stage 4 follow-up — cron overlap protection.
+--
+-- Every-minute analyses tick + narrative-LLM latency could leave a single
+-- analysis still running when the next tick fires. Without a claim, both
+-- ticks pick up the same row (last_run_at hasn't moved yet on success-only
+-- writes), and we double-spend on LLM calls and write duplicate result
+-- rows.
+--
+-- last_started_at is the claim timestamp; the runner takes a 90-second
+-- lease on the row before processing. A lease that's older than the
+-- threshold is considered abandoned and the next tick may re-claim.
+-- last_run_at continues to track completion time.
+
+ALTER TABLE analysis_definitions ADD COLUMN last_started_at TEXT;
+
+CREATE INDEX IF NOT EXISTS idx_analysis_definitions_started
+  ON analysis_definitions (project_id, last_started_at)
+  WHERE last_started_at IS NOT NULL;