Data model

The Postgres schema is the source of truth for everything Eden stores. This page walks the major tables and how they relate.

Core tables

telemetry_events

Every event from every SDK lands here. One row per event.

Column	Type	Notes
event_id	text	evt_<sha256[:16]> — content-derived
org_id	uuid	RLS-scoped
trace_id	text	trc_<sha256[:16]>
span_id	text	spn_<sha256[:16]>
parent_span_id	text	nullable
source_format	text	openai, anthropic, langchain, custom, etc.
name	text	e.g. openai.chat.completions, checkout
model	text	nullable
outcome	text	success, error, timeout
started_at	timestamptz
ended_at	timestamptz
duration_ms	integer
input_tokens	integer	nullable
output_tokens	integer	nullable
cost_usd	numeric(12,8)	nullable
prompt	text	PII-redacted
response	text	PII-redacted
attributes	jsonb	free-form
tags	text[]
created_at	timestamptz	defaults to now()

Indexes: (org_id, started_at DESC), (org_id, trace_id), (org_id, agent_id, started_at DESC), (org_id, model, started_at DESC).

judge_scores

One row per judge scoring, joined to telemetry_events on trace_id.

Column	Type
score_id	text (sc_<sha256[:16]>)
org_id	uuid
judge_id	text
trace_id	text
span_id	text
score	numeric(5,2)
passed	boolean
rubric	jsonb
reasoning	text
created_at	timestamptz

usage_events

Token and cost rollups — the source for cost intelligence.

Column	Type
usage_id	text (usg_<sha256[:16]>)
org_id	uuid
agent_id	text
model	text
event_type	text
quantity	integer
cost_usd	numeric(12,8)
attributes	jsonb
created_at	timestamptz

agents

Agent registry — one row per registered agent.

Column	Type
agent_id	text
org_id	uuid
name	text
framework	text
version	text
repo_url	text
tags	text[]
metadata	jsonb
registered_at	timestamptz
last_seen_at	timestamptz
archived_at	timestamptz

notifications

Alerts, in-app notifications, and the alert queue. Same table, discriminated by type.

Column	Type
id	uuid
org_id	uuid
user_id	uuid
type	text
title	text
body	text
data	jsonb
read_at	timestamptz
created_at	timestamptz

org_audit_log

Append-only audit log.

Column	Type
id	uuid
org_id	uuid
user_id	uuid
user_email	text
action	text
target_type	text
target_id	text
metadata	jsonb
ip_address	inet
user_agent	text
created_at	timestamptz

pii_redactions

Compliance evidence pack for SOC 2 / HIPAA / GDPR.

Column	Type
redaction_id	uuid
org_id	uuid
event_id	text
kind	text
field	text
created_at	timestamptz

Intelligence tables

Each intelligence service has its own table set. The full list:

Service	Tables
Trace replay	trace_replays, trace_replay_steps
RCA & incidents	incident_clusters, incident_alerts, incident_summaries
Cost intelligence	cost_summaries_daily, cost_anomalies, cost_forecasts
Retrieval intelligence	retrieval_logs, retrieval_gaps, chunk_attributions
Memory intelligence	memory_attributions, memory_timeline, memory_impacts
Agent intelligence	agent_performance_daily, agent_relationships, handoff_log
User intelligence	user_sessions, user_behavioural_signals, user_satisfaction_corrected
Governance	compliance_evidence_packs, data_lineage, pii_policies
Outcome / ROI	outcomes, outcome_rollups
Experimentation	experiments, experiment_splits, experiment_assignments, experiment_results
Predictive	predictions, drift_alerts
Autonomous	investigations, recommendations, self_heal_proposals
APM depth	apm_sampling_decisions, apm_k8s_enrichment_log, trace_step_profile_links

The full schema is in src/eden/db_migrations/versions/. Each migration is reversible (Alembic with explicit downgrade()).

Content-derived IDs

Almost every ID in Eden is content-derived:

event_id = "evt_" + sha256(org_id + trace_id + span_id + source_format + prompt + response + started_at)[:16]

This means re-sending the same payload produces the same event_id, so retries are idempotent at the storage layer. The downside is that you can't construct an ID independently — you have to hash the content. The SDKs and API do this for you.

Cardinality and growth

A mid-size customer (10M LLM calls/month) produces:

• ~10M rows in telemetry_events (300 GB / month at full prompt and response sizes; ~30 GB after compression and prompt truncation)
• ~3M rows in judge_scores (sampled)
• ~10M rows in usage_events

The retention policy (configurable per org) drops prompt and response after 30 days by default, keeping the metadata. This keeps the table size manageable without losing the ability to investigate historical regressions (which only need the attributes and judge_scores).

Migrations

Alembic. Every migration has a corresponding downgrade(). New migrations go through:

alembic revision --autogenerate -m "add foo table"
alembic upgrade head   # test locally
alembic downgrade -1   # verify the downgrade works
alembic upgrade head   # re-apply

Migrations run as a one-shot job in the deploy pipeline.