easy-db-lab

CH 01 · Validation for AI-built software

AI writes the code. Who proves it works?

Tests pass. Bugs still ship. easy-db-lab gives your agents a real cluster, real load, and real telemetry. They come back with a report, and every claim in it points to its evidence.

/plugin marketplace add rustyrazorblade/skills
/plugin install easy-db-lab@rustyrazorblade-plugins
/easy-db-lab:explore
Get startedSee a real report soon
Example session · Claude Code
you ▸ /easy-db-lab:plan validate my compaction branch
       against stock 5.0 under mixed load
agent ▸ wrote plan.md: 6 nodes, 2 arms, 3 workloads

you ▸ /easy-db-lab:run plan.md
agent ▸ easy-db-lab init lab-0412 --db.count 6 --up   ✓
        start cassandra: db0-2 stock, db3-5 branch  ✓
        stress: mixed read/write, 60 min            ✓
        pyroscope: wall-clock profile, db3          ✓
        report upload, then down                    ✓

verdict ▸ branch read p99 tbd vs stock tbd
          defects found in stock: tbd
          each claim checked against 2-evidence.md
The agent plans, runs, journals, and reports. You read the verdict first.

CH 02 · The missing phase

Most teams stop at tests.

Tests check what you thought of. A real cluster under real load shows what you did not think of: tail latency, GC pauses, disk limits, and the bug that appears only on node four after an hour.

  1. PromptAn agent gets the task.
  2. CodeThe agent writes it.
  3. TestsUnit and integration tests pass.
  4. Validateeasy-db-lab: real cluster, real load, profiles, and a report with evidence.
  5. MergeYou merge with proof.

CH 03 · How it works

Plan. Run. Read the verdict.

  1. 01

    Plan

    /easy-db-lab:plan

    Describe the test in plain words. The agent writes a plan with the cluster, the workloads, and a pass condition for each step.

  2. 02

    Run

    /easy-db-lab:run

    The agent provisions AWS, starts the database, drives load, and journals every command and observation as it goes.

  3. 03

    Report

    make -C docs

    Verdict first, evidence second: Grafana renders, stats tables, and raw CSV. Each summary claim is checked against its evidence before export.

CH 04 · Optimize

Then make it faster. Databases or your own app.

“Look at Pyroscope. Read CPU and wall-clock time. Make it faster.”

The agent uses the profiles as its feedback loop. Pyroscope, async-profiler, and eBPF see every JVM and every process the lab runs, so the loop works for Cassandra, ClickHouse, Spark, and your own service.

  1. load
  2. profile
  3. change
  4. measure
Beforewall-clock · compaction thread · illustration
Aftersame workload, same window
read p99
tbd
largest GC pause
tbd
CPU per GB compacted
tbd

CH 05 · Databases

Bring the database. Or add your own.

  • Apache Cassandra2.2 to trunk, mixed versions, your own branch
  • ClickHouseSharded, replicated, S3 storage
  • OpenSearchSearch and analytics
  • Apache SparkEMR clusters for analytics
  • PostgreSQLSysbench workloads
  • TiDBDistributed SQL
  • Trino and PrestoQuery engines
  • Apache KafkaStreaming
  • Neo4jGraph
  • Apache Ignite 3In-memory
  • Your databaseAn agent writes the kit:/easy-db-lab:create-kit valkey

CH 06 · Observability built in

Every node is instrumented before the first query.

CH 07 · Proof

Used by Fortune 500 engineering teams and Apache Cassandra committers.

Cursor compaction in Apache Cassandra

The cursor compaction, read, and flush work was built with agents and validated on easy-db-lab, side by side against stock Cassandra in one cluster.

Compaction allocation per byte
tbd
CPU per GB compacted
tbd
Read p99 under load
tbd

Advice you can trust

The cassandra-expert plugin gets its recommendations from easy-db-lab runs. Each one was tested on a real cluster, so your agent gives tested advice, not guesses from training data.

/plugin install cassandra-expert@rustyrazorblade-plugins

CH 08 · For engineering leaders

Want help with your infrastructure?

Rustyrazorblade Consulting tunes, validates, and fixes database infrastructure for large teams, with the same tools.

Work with Jon Haddad →