gem-hunter

Walk through PostgreSQL. Break things. Understand why.

An explorable 3D city where buildings represent PostgreSQL internals and motion shows their interaction. Follow a query, investigate a growing table, or see what changes when memory, checkpoints and replication become bottlenecks.

Explore the city → · Diagnose · Machine

Community discussion · Press coverage · Start here · Model accuracy

No installation. Runs in a browser with WebGL2. The city is an educational model, not a running PostgreSQL server.

930 points · 91 comments on Hacker News →

Explore the community’s questions, technical discussion and feedback on PGSimCity — How PostgreSQL Works, posted July 27, 2026.

Counts checked October 11, 2026. The discussion describes the version available then.

Selected reporting and hands-on reviews of PGSimCity:

Articles describe the version available when published.

A table keeps growing even though autovacuum is running. Why?

1. Open the city and choose Investigate. This starts the guided vacuum-blockade case.
2. Inspect and record the table, worker, snapshot and application-owner evidence. Explain what prevents cleanup before choosing an intervention.
3. End the transaction that the case establishes as abandoned, then check whether vacuum actually reclaims row versions. Releasing a snapshot is not itself cleanup.
4. Choose Inspect a page and its row versions to connect the investigation to the page layout, tuple header and snapshot diagrams.

Pause when you need time to read. Try challenge mode afterward for less guidance. Evidence and notes belong to the current attempt; they do not survive a reload.

The city is a model, not a running PostgreSQL server. Its numbers and timing are scaled to make internals observable. The separate Machine and opt-in Query flow can execute real PostgreSQL through PGlite; they do not turn the city into a production monitor.

The project is an evolving 0.x prototype, with known simplifications and tests against PostgreSQL behavior. See model accuracy and limitations for the reviewed reference version, formulas and review history.

PGSimCity targets the PostgreSQL 18 major line, with PostgreSQL 18.6 as its reviewed reference and REL_18_STABLE as the source branch. Four review rounds are documented in the accuracy notes; they do not establish complete coverage or make this model an emulator. Report a PostgreSQL mismatch.

Lessons include keyboard and text-first routes; the 3D scene and first-person walk do not have a nonvisual equivalent. Touch verification has used browser emulation, not physical devices. Accessibility and alternatives.

PostgreSQL 18's bulk-read strategy grows with io_combine_limit × effective_io_concurrency, subject to caps. This city's animation uses a fixed 32-frame ring instead; it is a historical simplification, not PostgreSQL 18's sizing rule. Model accuracy and source reference.

Colour is semantic everywhere and never decorative: WAL is amber, dirty pages are red, clean pages are blue, vacuum is violet, checkpoints are pink, the background writer is teal, replication is orange, storage is green, indexes are aqua, locks are red.

• Press T for the 14-chapter guided tour. It follows one connection from the client through planning, caching, WAL, checkpoints, vacuum and replication.
• Press Enter to trace one statement. Pick Non-HOT UPDATE and slow playback exposes where it enters the buffer pool, creates WAL and waits to commit.
• Run Cache thrash from the Scenarios menu. It sets shared_buffers to 16 MiB — below the manual control's 128 MiB minimum — so the clock sweep races and backends write their own dirty victims before they can read another page.
• Run The work_mem cliff. Its fixed Sort and HashAggregate nodes spill at 2 MiB, then fit at 4 MiB without replanning; the private reservoirs, base/pgsql_tmp, temp counters and latency breakdown show the consequence.
• Turn on Long-running transaction. The xmin horizon blade sinks and goes red; autovacuum still travels to the tables, but reports zero removable rows while the sessions table keeps bloating. Release the transaction and cleanup can begin again.
• Run Checkpoint storm. Watch the checkpointer's flywheel spin up, the fsync phase shudder, and a wall of full-page writes flood the WAL district after each checkpoint begins.
• Set synchronous_commit to off and watch backends stop waiting in commit_wait. Then read what you just traded away.
• Turn on Slow replay and watch sent_lsn, write_lsn, flush_lsn and replay_lsn pull apart on the standby.
• Press G and walk through the city at eye level. A buffer frame that read as one tile from the establishing shot becomes a structure above your head. Press E at the autovacuum lever or the postmaster door to operate it; the control center inside can trace one of six statements across a map of the city.
• Try an operator scenario, wait for its decision, and choose a response. Slot pressure, failover, and recovery make the consequence visible and offer a safe reset.

Start with drag to pan, wheel/pinch to zoom, T for the tour, K to pause and H to return to the overview.

Press ? in the city for the city control map and colour legend.

Choose Campaign, or open the required-standby case and the retired-consumer challenge. Opening preserves the city; Reset city and start explicitly creates a fresh case and clears replay history. Record ownership and live pressure, intervene, then observe and verify recovery. Explain why the same slot operation has a different trade-off when the consumer must return versus when it is retired.

Observation advances at most 300 model seconds per request and can be stopped. Slot removal releases one retention requirement; physical WAL reclamation waits for checkpoint completion and other retainers. Completion records an observed model result, not a graded explanation or measured learning gain. Only bounded case-completion identifiers persist locally; notes/evidence stay in this attempt and are discarded when switching cases. Clear local completion history removes that record. Storage failure leaves session-only progress usable.

Choose Replay, then Save checkpoint before an intervention. Observe an outcome, Rewind to checkpoint, change a model control or operator decision, and Run alternative to same duration. The comparison preserves the seeded prefix and labels counters, current gauges and model-time latency explicitly. Rewind reconstructs the model; saved investigation evidence and Diagnose’s previous-step button are navigation, not rewind.

The linked Console and Back to city controls transfer the current model incident in the same tab, including selection and diagnostic trail. Diagnostic sampling windows restart on arrival; PGlite remains a separate opt-in database. Export a comparison before switching views: saved replay checkpoints, comparison baselines and investigation notes do not transfer. Browser reload/Back is not linked navigation and can reject an expired or already-consumed transfer. Use the explicit new-city link to discard it. Replay records are local, bounded and exact-build-specific; copying a record is not a URL share, and does not include lesson answers or PGlite data.

Choose Export image in the city toolbar. The model pauses until you return to the city. PNG export keeps the current camera and graphics quality; visible object names are optional, but the model disclosure, version, and model clock always remain in the image footer. After a replay comparison, include the comparison explanation to retain the seed/checkpoint, recorded interventions, original and alternative durations, and qualified outcomes. Unequal durations are marked as not a controlled comparison. Long action lists are summarized with an explicit omitted count.

The default 2× option caps the scene at 4 megapixels. An explicit 4K scene option fits the camera aspect within 3840 × 2160 (rotated for a portrait view), subject to GPU limits. The disclosure footer adds height. Large captures can exhaust GPU memory; try 1× if an export fails. A prepared-image link remains available if the browser blocks the automatic download.

src/
  core/           shared contracts, event bus, registry, themes and utilities
  sim/            the PostgreSQL simulation
  world/          the city geometry, one module per district
  engine/         renderer, camera, flows, labels, picking, collision and audio
  ui/             controls, inspector, tour, search and written explanations
  observability/  a separate diagnostic interface over the same simulation
machine/           a separate psql workbench and 2D architecture board

Three rules hold it together:

1. world/layout.ts is the single source of truth for geography. Anchors, table definitions and the route network live there. No district hard-codes a coordinate another district needs.
2. The simulation never imports three.js, and the world never mutates the simulation. They meet at SimState.
3. Rendering carries meaning differently by theme. At night structure is matte and meaning is neon; in daylight hue and value carry meaning without relying on bloom. Local-time light follows an approximate 06:00–18:00 path from the reader's clock; it uses neither geolocation nor an astronomical latitude/season model.

Stack: three.js r185, TypeScript, Vite. three.js is the 3D application's only bundled runtime dependency. The separate 2D Query flow and Machine may lazy-load PGlite after reader opt-in. There is no framework, and Plausible analytics is the sole external service.

window.PGSIMCITY in the browser console includes sim, registry, bus, rig, gfx and flows if you would rather drive the city from the outside. For formulas, review history and known simplifications, see Model accuracy and limitations. Each inspector names material simplifications at the point where they matter.

The accuracy boundary makes internals such as the clock sweep's frame-by-frame victim choice observable. The separate Query flow and the Machine offer opt-in PGlite modes: real PostgreSQL supplies parsing, plans, catalogs, buffer counters, errors and results, while the visual model supplies the otherwise hidden interior. Each surface labels those sources separately because PostgreSQL exposes the former and not the latter.

You need Node.js ^20.19.0 || >=22.12.0 and a browser with WebGL2.

npm install
npm run dev      # http://localhost:5173

npm test
npm run typecheck
npm run build    # static bundle in dist/
npm run preview  # http://localhost:4173

There is no application server. The result is a static bundle. The 3D city and Diagnose model path make only the analytics requests described below. Query flow and the Machine may, after an explicit click or first submitted query, load the same-origin PGlite JavaScript, data and WebAssembly assets and run an in-memory PostgreSQL in the browser. Their model paths continue to work when analytics or PGlite is blocked.

Analytics and privacy. PGSimCity uses Plausible for aggregate, cookie-free analytics on the city, observability, and Machine pages. It records pageviews, unique visitors, referring sites, bounce rate, visit duration and interactions such as starting the tour, changing playback, opening a panel, tracing a statement, selecting a building or following an outbound link. PGSimCity sends no names, email addresses, free-form input, browser fingerprint or application-supplied personal data, and creates no analytics cookies, analytics local storage, advertising identifier or session recording. Blocking plausible.io stops measurement without affecting the application.

Follow the living delivery roadmap for current milestones and the technical roadmap for longer-term direction. See releases for what is actually shipped.

PGSimCity is an independent, non-commercial educational visualization of PostgreSQL internals. It is not affiliated with, sponsored, endorsed, or approved by Electronic Arts Inc. SimCity is a trademark of Electronic Arts Inc. This project contains no SimCity code, assets, artwork, logos, characters, audio, or game content.

Apache-2.0. Copyright 2026 Nikolay Samokhvalov. See NOTICE.

PostgreSQL is a trademark of the PostgreSQL Community Association of Canada. PGSimCity is an independent educational project and is not affiliated with, sponsored by, or endorsed by the PostgreSQL project.