Penguin Statistics

Penguin Statistics - V3 Backend

The re-designed Penguin Statistics v3 Backend, built with Go.

Using the data

The drop data is public:

• REST API at https://penguin-stats.io/PenguinStats/api/v2/: see the developer docs and the OpenAPI (Swagger 2.0) spec.
• MCP server for AI assistants (Claude, ChatGPT, Cursor, VS Code, Cherry Studio and others) at https://mcp.penguin-stats.io/mcp: Streamable HTTP, no authentication, read-only. In Claude Code:

  claude mcp add --transport http --scope user penguin-stats https://mcp.penguin-stats.io/mcp

The steps for every other app, in English and Chinese, are at https://mcp.penguin-stats.io/.

• Plugin and skill for coding agents, the MCP server with a skill on reading, citing and querying the numbers: https://github.com/penguin-statistics/agent-skills

The drop data is licensed CC BY-NC 4.0: non-commercial use, with credit. Credit it as Data: Penguin Statistics (penguin-stats.io), CC BY-NC 4.0. The code of this repository is MIT.

Architecture

This Go project mainly uses:

• gofiber/fiber
• uber-go/fx
• uptrace/bun

This project mainly follows a RESTful API design.

Operations

Report tasks in the dead-letter stream

The report worker moves a report task out of the penguin-reports stream into penguin-reports-dead (subjects REPORT_DEAD.*) when it can never be processed (it is not a task, or a stage of it does not exist) or when it failed at each of its 30 deliveries, which take about 97 minutes of retry delays. Each move is logged at error level (reportwkr.malformed, reportwkr.parked or reportwkr.dead). The copy keeps the payload as published, with these headers:

• Penguin-Original-Subject: the subject it was published under, like REPORT.SINGLE
• Penguin-Original-Sequence: its sequence in penguin-reports
• Penguin-Deliveries: its deliveries when it moved
• Penguin-Last-Error: the error that moved it

A task too close to the maximum payload to fit with these headers moves without them, and its subject under REPORT_DEAD.* names its original subject; its move is logged as reportwkr.dead.bare, with the sequence and the error.

Nothing in the stream expires. With the nats CLI, once the cause is fixed:

# list the tasks, with their headers

nats stream view penguin-reports-dead

show one, by its sequence in penguin-reports-dead

nats stream get penguin-reports-dead 42

replay it under its original subject: the worker processes it as a new task

nats stream get penguin-reports-dead 42 --json | jq -r .data | base64 -d \ | nats pub REPORT.SINGLE --force-stdin

remove it once replayed

nats stream rmm penguin-reports-dead 42 -f

A task that moved after failing at each delivery may have been stored by one of them whose commit outcome was unknown (for example the connection dropped during the commit); check drop_reports for its account and time before replaying it.

Benchmarks

Report consumer CPU

internal/workers/reportwkr/consumerbench_test.go (build tag consumerbench) drives the report consumer through its real code path: each task is published to a JetStream stream of its own, fetched and handed to Worker.ingestPreprocess (BenchmarkConsumer/ingest), which stores it in PostgreSQL and sets its recall key in Redis, with logging and tracing configured as in production. BenchmarkConsumer/process runs Worker.process alone. Besides wall time and allocations it reports, per task, the CPU time of the process (cpu-ns/op, user + system from getrusage, which a shared machine disturbs far less than the wall time) and the time spent waiting on PostgreSQL (pg-wait-ns/op, pg-wait-%) and Redis.

The tasks (internal/workers/reportwkr/testdata/consumerbench/tasks.jsonl.gz) are encoded as the report handler publishes them, drawn from the stages, servers, times, sources, versions and drop patterns of recent production reports, with synthetic accounts, IPs and task ids.

It needs a scratch database, Redis and NATS with JetStream. The scratch database holds the schema and the reference tables of a restore of production (penguinprod, only read) and synthetic accounts; the benchmark truncates its drop_reports and drop_report_extras:

scripts/consumerbench/setup-db.sh   # creates penguin_cpubench and the role cpubench on 127.0.0.1:5434

regenerate the tasks, if needed: CONSUMERBENCH_TASKS, CONSUMERBENCH_BATCH_PERMILLE, CONSUMERBENCH_SEED

CONSUMERBENCH_GENERATE=1 go test -tags consumerbench -run TestConsumerBenchGenerate ./internal/workers/reportwkr/

The defaults point at penguin_cpubench on 127.0.0.1:5434, Redis DB 14 on 127.0.0.1:16379 and NATS on 127.0.0.1:4222 (CONSUMERBENCH_POSTGRES_DSN, CONSUMERBENCH_REDIS_URL, CONSUMERBENCH_NATS_URL). Then:

# one run

go test -tags consumerbench -run '^$' -bench BenchmarkConsumer -benchtime 2000x ./internal/workers/reportwkr/

interleaved runs of variants, compared with benchstat: logging to a counting io.Discard (prod), to /dev/null

(devnull), disabled (logoff); tracing as in production at 10% (prod), off (traceoff), at 100% (trace1); both off (off)

COUNT=10 scripts/consumerbench/run.sh bench prod logoff traceoff off devnull trace1

CPU and allocation profiles of the ingestion, the top functions and the share of each part of the consumer

scripts/consumerbench/run.sh profile prod

Outputs go to $OUT (default a new temporary directory). Each run logs the load average and the queries other sessions had active on the cluster to $OUT/runs.log. The PostgreSQL connection runs over loopback, where a write system call also does the receiving side's TCP work, so the system calls weigh more than against a remote server.

Maintainers

This project has mainly being maintained by the following contributors (in alphabetical order):

• AlvISsReimu
• GalvinGao

The full list of active contributors of the _Penguin Statistics_ project can be found at the Team Members page of the website.

Contributors

This project is made possible by the following contributors. Contributions are always welcome!

License