Keep the events
What the brief asks, how large the traffic is, and the architecture I would ship. Demo boards at the end. Numbers marked brief, public, or estimate.
Task
Review the CrazyGames analytics pipeline and say how to improve it: monitoring, closer-to-real-time streaming, and company-wide LLM access with cost and permission control.
What I send
A written report and a small running demo of the same hops. Airflow every 20 minutes, Postgres as the warehouse stand-in, two Grafana boards. Synthetic events only.
What the brief forbids
A report that is mostly generated text with no decisions. AI as a daily tool is fine. The architecture choices and the reasons for them have to be mine.
Problem
About 4.000.000 daily players feed product, recommendations, developer stats, and ads through one event path. Five facts matter.
- LossAccepted events can vanish. ECS accepts the event, holds it in memory by type, then writes JSON to S3. A deploy or OOM before the flush loses the batch. The Redshift job stays green because there was never a file.
- LagFreshness is a chain. Play, memory, S3, copy every 20 minutes, hourly or daily dbt, then the portal. A developer watching plays after a release feels that whole chain. Daily training does not.
- LoadOne warehouse is shared. Raw load, dbt, Preset, stats, userportal, and model-cron compete for Redshift Serverless RPU.
- AccessClaude without a gate. Company-wide SQL through an LLM needs identity, row filters, a metric catalog, and a hard cost stop. A shared warehouse password is a leak.
- KeepAlready right. One endpoint for four clients. Validation before storage. S3 as durable handoff. Serverless for scheduled COPY. dbt as the single definition of a play.
Data
Scale from the brief, their job posts, AWS, and company press.
| Metric | Value | Source |
|---|---|---|
| Daily active users | 4.000.000 | Test brief |
| Daily active users | 3.000.000+ | Product Analyst job |
| Monthly players | 50 to 60.000.000 | Cloudflare / company posts |
| Games on platform | 3.000+ | AWS Games Blog 2024 |
| Gameplays / month | ~300.000.000 | Data Engineer job |
| Gameplays in 2026 | 3.000.000.000 | Company press Oct 2026 |
| Playtime / day | > 2.000.000 h | Product Analyst job |
| Playtime YTD to Sep 2026 | 700.000.000 h | CPO public post |
| Ingest average (estimate) | ~1.000 to 2.300 / s | From event estimate |
| Ingest peak (estimate) | ~3.000 to 10.000 / s | About 3 to 5× average |
| Per 20-min window (estimate) | ~1 to 3M rows | 72 windows / day |
Event sources in the brief: portal, game frame (GF), SDK, ads. Server-side bot filter, integrity check, field validation, schema cleanup, enrichment.
Architecture today
As drawn in the brief. Batch warehouse, not a stream.
Pink border marks the silent loss hop. Everything after S3 can retry a file that exists.
Architecture proposed
Keep the batch spine. Fix the lossy hop. Add one narrow fast path. Put Claude behind a query tool.
Narrow stream
The same S3 flush feeds a small consumer for game_start and game_end into a stats table (about one minute). The developer portal reads that. Hourly dbt stays the correction. Upsert by event_id.
What I would not do
Replace the whole 20-minute COPY with a full stream of every event type. At this volume that buys duplicates, out-of-order repairs, and a larger bill, for screens nobody is watching live.
- RuleS3 before forget. Memory may drop only after JSON is on disk. Shutdown flushes.
- RuleStream one screen. Developer play counts need minutes. Recommendations and day aggregates stay on dbt.
- RuleOne definition of a play. The stream does not invent a second formula.
- RuleClaude never gets a warehouse password. The tool runs as the person’s read-only role, with a mart allowlist, a metric catalog, and an RPU hard stop.
Live demo
Same hops on our Airflow and Grafana. Synthetic traffic. Use it to show buffer loss, not production volume.
Trigger the DAG with {"crash_buffer": true} to drop one in-memory batch. Events lost in the buffer rises. Rows already loaded stay.