Website:
Job details:
Telemetry & TimescaleDB Architecture
This document tracks the design decisions and architecture for handling high-frequency health data (IoT devices, wearables, continuous monitors) in Health Assistant.
The Problem
Standard clinical data maps beautifully to the HL7 FHIR Observation model. A blood test taken once every 3 months is easily stored in the fhir_observations PostgreSQL table. However, devices like the Apple Watch, Oura Ring, or continuous glucose monitors (CGMs) can generate health measurements every 1 to 5 minutes. Routing this data into standard FHIR structures results in massive table bloat, drastically degrading dashboard performance and slowing down routine clinical queries.
The Solution: Split Architecture
Health Assistant Implements a Frequency-based Routing Architecture
- Low-Frequency (Clinical) Data -> FHIR Observations
- Stored in standard fhir_observations table.
- Used for blood panels, point-in-time weight, diagnostic results.
- High-Frequency (Telemetry) Data -> TimescaleDB
- Stored in the telemetry_data hypertable.
- Extremely high ingestion limits and heavily compressed.
Dynamic Routing via Biomarker Definitions
Instead Of Hardcoding Which Metrics Go To Which Database, Routing Is Configurable. The Biomarker_definitions Table Includes An Is_telemetry Boolean Flag. When Integration Webhooks Push Data To The Backend, The Parser Resolves The Data To a Standard Biomarker. The Core Sync Engine Then Inspects This Flag
- If is_telemetry = true: Data is mapped to a TelemetryDataModel and saved to TimescaleDB.
- If is_telemetry = false: Data is mapped to an Observation and saved to FHIR.
Automated Data Migration: System administrators can toggle this flag via the UI, making the system infinitely expandable for future IoT devices without requiring code changes (see "Long-Format Storage" below for why this is now truly zero-code). When the flag is toggled on an existing biomarker, the Celery task migrate_biomarker_data performs an automatic, batched (5000 rows) migration between the standard PostgreSQL fhir_observations table and the TimescaleDB telemetry_data hypertable.
Telemetry → FHIR patient attribution: every telemetry row carries a persisted patient_id column (populated at insert by the integration pipeline, the FHIR→telemetry migration direction, and the /telemetry/data upload endpoint). When an admin flips is_telemetry true→false on a populated biomarker, the migration task resolves the patient per row directly from tr.patient_id — no device-id attribution chain needed.
For rows that lack a persisted patient_id (e.g. legacy mobile uploads that didn't send one), the single-patient-tenant fallback remains the safe default. Rows that can't be attributed (no patient_id + multi-patient tenant) are
skipped, not silently assigned to a random patient. The count is exposed in BiomarkerDefinition.meta_data["migration_skipped_no_patient"].
Long-Format Storage
The telemetry_data hypertable is
long-format: one row per (timestamp, device, slug) triple with explicit value Float + unit Text columns. This is the canonical TimescaleDB / InfluxDB / Prometheus pattern for variable-schema metrics.
Column Type Notes tenant_id UUID No FK (TimescaleDB limitation); cleanup job purges on tenant delete. id UUID Composite PK part (required by TimescaleDB). timestamp timestamptz Composite PK part; hypertable time dimension. device_id String Source device / integration instance. slug String Biomarker slug (heart-rate, steps, spo2, glucose, …). value Float NOT NULL — the numeric measurement. unit String Optional unit symbol (bpm, count, %). patient_id UUID Persisted attribution (populated at insert).
Why long-format (vs. the legacy wide+JSONB hybrid with dedicated heart_rate/steps/calories columns):
- Adding a new telemetry biomarker is a row-only change. Flip BiomarkerDefinition.is_telemetry = True — no DDL, no service-layer branching, no new column. SpO2 / CGM / sleep stages are first-class citizens alongside heart rate.
- Generic continuous aggregates. One GROUP BY slug CAgg definition covers every current and future telemetry biomarker.
- Range queries get a real B-tree. WHERE value > 120 prunes via (tenant_id, slug, timestamp) instead of needing per-metric JSONB functional indexes.
The previous _METRIC_COLUMNS alias map, the slug→column branching in 5 files, and the JSONB data catch-all are gone. See migration t1e2l3o4n5g6_telemetry_long_format for the full rationale.
Preserving Spikes (OHLC Aggregation)
When rendering a dashboard for "The Last 6 Months", fetching minute-by-minute heart rate data would crash a user's browser. Data must be downsampled. However, taking a pure mathematical AVERAGE() hides critical clinical data (e.g., a dangerous 10-minute spike to 190 bpm).
To solve this, the AnalyticsService uses
TimescaleDB OHLC (Open-High-Low-Close) Aggregations:
- time_bucket_gapfill() ensures missing days (e.g. user forgot to wear the watch) are returned as NULL rather than connected across the graph.
- The query returns AVG(), MIN(), and MAX() for each bucket.
- The frontend renders this as an average line with a shaded variance envelope (Min to Max), guaranteeing that the absolute highest spike during that timeframe remains visibly represented on the chart.
Decoupled Aggregation Resolution: The system decouples the "Temporal Scope" (e.g., viewing the Last 30 Days) from the "Aggregation Bucket" (e.g., grouping by 1-hour averages). Users can dynamically adjust the resolution density directly from the UI dropdowns, and the backend securely handles the dynamic PostgreSQL INTERVAL casting (e.g. 1 minute, 15 minutes, 1 day, 1 week).
When the requested bucket exactly matches a continuous aggregate's resolution, the analytics service transparently dispatches to the CAgg (pre-computed, fast on long horizons); otherwise it falls back to live time_bucket_gapfill() over the raw hypertable. The CAgg path filters WHERE slug = :slug so every telemetry biomarker is covered by one definition.
Implemented optimizations (TimescaleDB)
The Hypertable Is Paired With Production-grade TimescaleDB Features, Both Registered In The Consolidated Baseline (alembic/versions/8ddb7ef7ca4d_consolidated_baseline.py) And Rebuilt Long-format In Migration T1e2l3o4n5g6_telemetry_long_format
- Generic continuous aggregates: three materialized views — telemetry_hourly, telemetry_daily, telemetry_monthly — pre-compute rollups in the background via add_continuous_aggregate_policy. Each groups by slug, so one definition covers every current and future telemetry biomarker (no DDL when a new biomarker is flagged is_telemetry=True). The monthly CAgg covers the last-12-months / all-time analytics buckets that previously hit the raw hypertable. The AnalyticsService (backend/app/services/analytics_service.py) transparently dispatches to these CAggs when the requested bucket is 1 hour, 1 day, or 1 month.
- Compression policy: add_compression_policy('telemetry_data', INTERVAL '7 days') compresses raw minute-by-minute data older than one week (compress_segmentby = 'tenant_id, device_id, slug', compress_orderby = 'timestamp DESC'). Segmenting by the query-pruning keys means both tenant-wide and per-device queries skip irrelevant compressed chunks.
- Retention policy: add_retention_policy('telemetry_data', INTERVAL '2 years') auto-prunes raw rows older than 2 years.
Upload Contract
The POST /api/v1/telemetry/data endpoint accepts long-format points — one per metric/timestamp:
{ "device_id": "apple_watch_1", "points": [ {"timestamp": "2026-07-28T08:00:00Z", "slug": "heart-rate", "value": 72.0, "unit": "bpm"}, {"timestamp": "2026-07-28T08:00:00Z", "slug": "steps", "value": 12.0}, {"timestamp": "2026-07-28T08:05:00Z", "slug": "heart-rate", "value": 75.0, "unit": "bpm"} ] }
A mobile client syncing N metrics at one timestamp sends N points (not one point with N fields). The integration SDK already produces one Observation per metric, so the mapping is 1:1. The service uses SQLAlchemy Core bulk insert with ON CONFLICT DO NOTHING chunked at 5000 rows for high-frequency ingest.
Future Considerations (Roadmap)
- FHIR Interoperability Boundary: The split architecture inherently moves high-frequency telemetry data outside of strict FHIR compliance. Currently, when exporting patient records to FHIR, telemetry data is excluded. Future versions will need to dynamically downsample and map TimescaleDB data back into FHIR Observation bundles during export.
- Continuous-aggregate-based anomaly detection: get_telemetry_anomalies still runs the detector in Python over raw rows; SQL-side downsampling via the CAggs is a separate improvement.
Unified Clinical View (UI & AI Integration)
Despite the data being physically split across two different database engines, both the frontend and the AI Chatbot provide a unified longitudinal view.
- Frontend: The BiomarkerDetail view uses the AnalyticsService to merge and sort FHIR and TimescaleDB data.
- AI Chatbot: The AI Assistant uses the get_aggregated_biomarker_trends tool, which routes through the same AnalyticsService logic. This ensures the AI can reason over high-frequency wearable data (steps, heart rate, SpO2) without being overwhelmed by raw records, while strictly adhering to aggregated OHLC (Average, Min, Max) values.
This is achieved dynamically by the AnalyticsService in the backend:
- FHIR Data: The service queries standard fhir_observations matching the biomarker_id.
- Telemetry Data: The service queries the long-format hypertable (or the matching continuous aggregate) with WHERE slug = :slug — uniform across every telemetry biomarker.
- Merge & Sort: The backend formats both datasets into the identical JSON response structure, sorts them chronologically by timestamp, and returns them to the requester (React chart or AI Tool) as a single cohesive longitudinal record.
Click on Apply to know more.