Phase 6: Analytics and Event Log Offload

Goal

Keep Prisma Postgres focused on transactional product state. Move high-volume analytics, logs, traces, raw payloads, and event streams out of Prisma before they distort operation count, storage, and performance.

Current-state audit

Schema search

Run:

grep -RniE "analytics|event|events|log|logs|audit|trace|metric|pageView|click|impression|webhook|usage|token|generation|scheduler" prisma src

Write-path search

Run:

grep -RniE "prisma\.[a-zA-Z0-9_]+\.(create|createMany|update|upsert)" src

Classify every write path.

DB table audit

SELECT
  schemaname,
  relname AS table_name,
  n_live_tup AS approx_rows,
  pg_size_pretty(pg_total_relation_size(relid)) AS total_size,
  pg_total_relation_size(relid) AS total_bytes
FROM pg_stat_user_tables
ORDER BY pg_total_relation_size(relid) DESC;

Event/log candidates:

SELECT
  relname AS table_name,
  n_live_tup AS approx_rows,
  pg_size_pretty(pg_total_relation_size(relid)) AS total_size
FROM pg_stat_user_tables
WHERE relname ILIKE '%event%'
   OR relname ILIKE '%log%'
   OR relname ILIKE '%audit%'
   OR relname ILIKE '%metric%'
   OR relname ILIKE '%trace%'
   OR relname ILIKE '%analytics%'
   OR relname ILIKE '%webhook%'
ORDER BY pg_total_relation_size(relid) DESC;

Classification rules

Data type	Keep in Prisma?	Destination
Transactional audit log visible in app	Maybe	Prisma, capped/archived
User-visible activity feed	Maybe	Prisma, capped/archived
Raw product analytics	No	PostHog
Page views / clicks	No	PostHog
Session replay	No	PostHog
AI traces / generation spans	No	Langfuse or PostHog AI Observability
Raw webhook payload archive	No	R2
Scheduler heartbeat logs	No	Axiom/GCP logs
API request logs	No	Axiom/GCP logs
Error logs	No	Axiom/GCP logs/PostHog error tracking
Customer-facing analytics API	Not primary Prisma	Tinybird/ClickHouse/BigQuery

Future TendSocial event categories

Category	Examples	Recommended destination
Product analytics	feature usage, page views, funnels	PostHog
UX/session replay	sessions, rage clicks, form friction	PostHog
App logs	API logs, worker logs, queue failures	Axiom or Google Cloud Logging
AI observability	prompt, model, latency, cost, traces	Langfuse or PostHog AI Observability
Raw payload archive	webhook payloads, social sync payloads	Cloudflare R2 NDJSON/Parquet
Customer analytics	content performance dashboards	Tinybird/ClickHouse/BigQuery
Long-term BI	revenue, cohort, customer reporting	BigQuery or ClickHouse

Destination options and pricing snapshot

PostHog

Best for product analytics, web analytics, feature flags, session replay, errors, and some AI observability.

Current free tier includes 1M analytics events/month, 5k session replay recordings, 1M feature flag requests, 100k exceptions, 1M data warehouse rows, 100k AI observability events, and 50 GB logs ingest. Product analytics overage starts at $0.00005/event and gets cheaper at higher volume. Source: PostHog pricing.

Cloudflare R2

Best for cheap raw append-only archive: webhook payloads, social sync payloads, NDJSON, and Parquet batches.

R2 Standard free tier includes 10 GB-month storage, 1M Class A operations, 10M Class B operations, and free egress. Standard storage example pricing implies $0.015/GB-month after the free tier. Source: Cloudflare R2 pricing.

Axiom

Best for application logs and operational telemetry.

Axiom Cloud's Personal plan has no charge and includes 500 GB/month data loading, 10 GB-hours/month query compute, and 25 GB/month storage. Axiom also describes Always Free allowances for Axiom Cloud: 1,000 GB/month data loading compute, 100 GB-hours/month query compute, and 100 GB/month storage. Source: Axiom pricing.

Tinybird

Best for custom real-time analytics APIs backed by ClickHouse.

Tinybird Free includes shared infrastructure, 0.25 vCPU, 1 thread/request, 1k requests/day, 10 GB included storage. Developer starts at $25/month. Source: Tinybird pricing.

BigQuery

Best for long-term warehouse and ad-hoc analytical queries.

Google Cloud Free Tier includes 1 TiB querying/month and 10 GiB storage/month for BigQuery. Source: Google Cloud Free Tier.

Langfuse

Best for LLM/AI traces, prompt evaluation, generation cost/latency tracking, and AI observability. Use only if AI trace debugging becomes important enough to justify a specialized tool. Source: Langfuse pricing.

Recommended starting architecture

Prisma Postgres
  transactional product state only

PostHog
  product analytics
  web analytics
  feature usage
  session replay if needed
  error tracking if desired

Cloudflare R2
  raw webhook payload archive
  raw social sync payload archive
  large append-only event exports

Axiom or Google Cloud Logging
  backend logs
  worker logs
  scheduler logs
  operational errors

Langfuse or PostHog AI Observability
  AI traces and generation observability

Tinybird/BigQuery later
  customer-facing analytics or warehouse reporting

Event producer abstraction

Create one internal interface so app code does not directly write analytics/log events to Prisma:

export const analytics = {
  track: async (eventName: string, properties: Record<string, unknown>) => {
    // PostHog
  },
};

export const eventArchive = {
  write: async (stream: string, payload: unknown) => {
    // R2 NDJSON/Parquet later
  },
};

export const appLog = {
  info: async (message: string, fields?: Record<string, unknown>) => {
    // Axiom/GCP logger
  },
  error: async (message: string, fields?: Record<string, unknown>) => {
    // Axiom/GCP logger
  },
};

export const aiTrace = {
  record: async (payload: unknown) => {
    // Langfuse/PostHog AI Observability
  },
};

Migration tasks

1. Identify current Prisma event/log/analytics tables.
2. Identify current write paths.
3. Classify each table/write path.
4. Keep only transactional records in Prisma.
5. Stop high-volume writes to Prisma.
6. Route product analytics to PostHog.
7. Route raw payload archive to R2.
8. Route operational logs to Axiom/GCP.
9. Route AI traces to Langfuse/PostHog if needed.
10. Backfill/export existing data only if useful.
11. Drop or truncate old high-volume tables after validation.
12. Add code review guardrail: no new high-volume event tables in Prisma without explicit approval.

Acceptance criteria

Item	Target
Current analytics/event/log tables identified	100%
Current high-volume write paths identified	100%
Destination chosen for each stream	Yes
New high-volume writes removed from Prisma	Yes
Existing data archived/exported where useful	Yes
Prisma contains only transactional product state	Yes
Operation tests re-run after offload	Yes
Prisma monthly operation estimate recalculated	Yes

Sources

• PostHog pricing
• Cloudflare R2 pricing
• Axiom pricing
• Tinybird pricing
• Google Cloud Free Tier
• Langfuse pricing