Airtable Engine is a high-performance Airtable-style spreadsheet built to keep tables with millions of rows responsive, achieving sub-millisecond read latency and millisecond row-write latency at million-row scale. The browser holds only a small virtualized window while the server resolves every scroll position through a three-tier read path built on indexed keyset seeks, per-view row rank tables, JSONB rows, and cursor-anchored jumps.
Live: airtable-engine.vercel.app (Google sign-in)
The bulk-add control is the quickest way to create a large table and try a deep scrollbar jump.
Below are median latencies across 15 runs at one million records, for the various paths the grid depends on.
| Operation | Median latency |
|---|---|
| Jump into an unsorted table (Tier 1, keyset seek) | 0.2 ms |
| Jump into a saved sorted view (Tier 2, precomputed ranks) | 1.2 ms |
| Jump with an unsaved ad-hoc sort (Tier 3, cursor anchor) | 48.9 ms |
| Jump into a filtered view (Tier 3, cursor anchor) | 76.9 ms |
| Scroll one page (keyset) | 0.2 ms |
| Search, first page of matches | 15.3 ms |
For contrast, adding a cursor anchor reduces an ad-hoc sorted jump from 3,569 ms to 48.9 ms, a 73x improvement. At the same depth, Tier 1's indexed keyset seek completes an unsorted jump in 0.2 ms, compared with 103 ms using naive OFFSET pagination, a 515x improvement. The read tiers are explained in scaling engine, with p95 and the 1K and 100K comparisons in full benchmark results.
Product walkthrough
Bases can be starred, and recently opened work stays immediately accessible from the dashboard.
The grid fetches and renders only the visible window while preserving direct access to all 1,000,025 records.
Filter groups combine text and number conditions through nested AND/OR logic, with matching rows updated in place.
Each view retains its own configuration, while saved sorts are committed for rank-backed jumps through the result set.
Find-in-view highlights the active result, reports the full match count, and navigates to matches outside the loaded window.
Keyboard-driven edits appear in the local grid immediately while the mutation completes in the background.
You need Node 20+, pnpm, Docker, and a Google OAuth client.
# Install dependencies
pnpm install
# Configure the app
cp .env.example .env
# Start PostgreSQL and apply the schema
./start-database.sh
pnpm prisma migrate dev
# Start the app
pnpm devstart-database.sh starts PostgreSQL in Docker. The image tag is currently unpinned; if you already have Postgres, use that connection in .env instead. Google OAuth setup and the required variables are covered in deployment.md.
| Feature | Engineering detail |
|---|---|
| Deep scrolling | A custom scrollbar maps pixel offset to a row index and fetches only that window on a composite index, with visited windows cached. |
| Two sort modes | A saved sort precomputes a per-view rank table for indexed jumps, and an unsaved ad-hoc sort anchors its jumps to a cursor over a JSONB expression index. |
| Field duplication | The new column points at its source field, so it is readable, sortable, and filterable while a batched backfill copies the JSONB keys behind it. |
| Row order | Rows carry fractional positions, so an insert or a drag is one midpoint write instead of a renumber. |
| Field types | Text and number fields share one JSONB cell store, and sorting either builds a typed expression index, collated for text and cast to double precision for numbers, so numbers sort by value rather than as text. Number fields add per-field decimal places, separators, and abbreviation, so 3456 shows as 3.5K and parses back on input. |
| Filters | Nested AND/OR groups compiled into a single parameterized query, anchored the same way as sorted views. |
| Search | Substring matching across all cells, with a live match count and previous or next navigation that resolves off-window matches server-side. |
| Bases and views | Bases group tables and surface starred and recently opened ones first, and each table carries multiple views whose filters, sort, column order, hidden fields, and row height all persist independently. |
| Cell editing | Keyboard-driven mutations optimistically patch both paginated query state and the jump cache. Failed writes restore paginated state, while cached jump windows refresh on the next navigation. |
| Interface | Every screen is a 1:1 recreation of Airtable's, matched by overlaying the original and hand-written with no component library underneath. |
Start with Architecture for the big picture, then Scaling engine for the read and write paths.
| Document | What it covers |
|---|---|
| Architecture | System layers, tiers, stack, and repository |
| Scaling engine | Read and write paths with benchmark results |
| Data model | Prisma schema, JSONB cells, ordering, ranks |
| API reference | tRPC routers and procedures |
| Deployment and configuration | Environment variables, limits, and hosting |
This project is licensed under the PolyForm Noncommercial License 1.0.0.
Origin. This project was originally built as a solo technical assessment for Lyra, based on a brief to build an Airtable clone.