|
| 1 | +# pg_column_tetris |
| 2 | + |
| 3 | +A PostgreSQL extension that enforces optimal column alignment to minimize row padding waste. |
| 4 | + |
| 5 | +## Why It Matters |
| 6 | + |
| 7 | +PostgreSQL aligns columns to their type's natural boundary (1, 2, 4, or 8 bytes). When columns are ordered poorly, padding bytes fill the gaps. These wasted bytes don't just cost disk space — they cost **memory**. Padding is loaded as-is into `shared_buffers` and OS page cache, meaning more bytes per row → fewer rows per 8 KB page → more cache pressure → more disk I/O. |
| 8 | + |
| 9 | +For a table with millions of rows, reordering columns can reclaim gigabytes of memory. |
| 10 | + |
| 11 | +## Installation |
| 12 | + |
| 13 | +```bash |
| 14 | +# From source |
| 15 | +make install |
| 16 | + |
| 17 | +# Then in PostgreSQL |
| 18 | +CREATE EXTENSION pg_column_tetris; |
| 19 | +``` |
| 20 | + |
| 21 | +No compilation needed — pure SQL/PL/pgSQL. Works on any PostgreSQL instance that supports event triggers (RDS, Cloud SQL, Supabase, Neon, etc.). |
| 22 | + |
| 23 | +## Quick Start |
| 24 | + |
| 25 | +The extension installs in **warn** mode by default. Create a table with suboptimal column order: |
| 26 | + |
| 27 | +```sql |
| 28 | +CREATE TABLE orders ( |
| 29 | + is_shipped boolean, |
| 30 | + order_total numeric, |
| 31 | + user_id bigint, |
| 32 | + item_ct smallint, |
| 33 | + order_dt timestamptz, |
| 34 | + status smallint, |
| 35 | + ship_dt timestamptz |
| 36 | +); |
| 37 | +``` |
| 38 | + |
| 39 | +You'll see a NOTICE suggesting the optimal order: |
| 40 | + |
| 41 | +``` |
| 42 | +NOTICE: pg_column_tetris: suboptimal column alignment — 19 bytes of fixed-width padding wasted per row |
| 43 | +``` |
| 44 | + |
| 45 | +Switch to **strict** mode to block suboptimal tables entirely: |
| 46 | + |
| 47 | +```sql |
| 48 | +SELECT pg_column_tetris.set_mode('strict'); |
| 49 | +``` |
| 50 | + |
| 51 | +Now the same CREATE TABLE will fail with an error and a hint showing the optimal column order: |
| 52 | + |
| 53 | +``` |
| 54 | +ERROR: suboptimal column alignment — 19 bytes of fixed-width padding wasted per row |
| 55 | +HINT: Suggested order: |
| 56 | + CREATE TABLE orders ( |
| 57 | + user_id bigint, -- 8-byte aligned |
| 58 | + order_dt timestamptz, -- 8-byte aligned |
| 59 | + ship_dt timestamptz, -- 8-byte aligned |
| 60 | + item_ct smallint, -- 2-byte aligned |
| 61 | + status smallint, -- 2-byte aligned |
| 62 | + is_shipped boolean, -- 1-byte aligned |
| 63 | + order_total numeric -- varlena (last) |
| 64 | + ); |
| 65 | +``` |
| 66 | + |
| 67 | +## Optimal Column Order |
| 68 | + |
| 69 | +The extension enforces this ordering rule for fixed-width columns: |
| 70 | + |
| 71 | +1. **8-byte aligned** (`d`): `bigint`, `timestamptz`, `float8`, `interval` |
| 72 | +2. **4-byte aligned** (`i`): `integer`, `float4`, `date`, `oid` |
| 73 | +3. **2-byte aligned** (`s`): `smallint` |
| 74 | +4. **1-byte aligned** (`c`): `boolean`, `char(1)` |
| 75 | +5. **Variable-length** (varlena): `text`, `varchar`, `numeric`, `jsonb`, `bytea` — always last |
| 76 | + |
| 77 | +Within each group, `NOT NULL` columns are preferred first (minor CPU optimization for tuple deforming). |
| 78 | + |
| 79 | +## Auditing Existing Tables |
| 80 | + |
| 81 | +```sql |
| 82 | +-- Detailed layout report for a single table |
| 83 | +SELECT * FROM pg_column_tetris.check('orders'); |
| 84 | + |
| 85 | +-- Generate migration DDL to fix a table |
| 86 | +SELECT pg_column_tetris.suggest_rewrite('orders'); |
| 87 | +``` |
| 88 | + |
| 89 | +The `suggest_rewrite` function generates a complete migration script: |
| 90 | + |
| 91 | +```sql |
| 92 | +BEGIN; |
| 93 | +ALTER TABLE public.orders RENAME TO orders_old; |
| 94 | +CREATE TABLE public.orders ( ...optimal order... ); |
| 95 | +INSERT INTO public.orders SELECT ... FROM public.orders_old; |
| 96 | +DROP TABLE public.orders_old; |
| 97 | +COMMIT; |
| 98 | +``` |
| 99 | + |
| 100 | +## Configuration |
| 101 | + |
| 102 | +```sql |
| 103 | +-- Set mode: 'strict' (block), 'warn' (notice only), 'off' (disable) |
| 104 | +SELECT pg_column_tetris.set_mode('strict'); |
| 105 | + |
| 106 | +-- Check current mode |
| 107 | +SELECT pg_column_tetris.mode(); |
| 108 | + |
| 109 | +-- Exclude a table from validation (e.g., matching an external schema) |
| 110 | +SELECT pg_column_tetris.exclude('legacy_imports'); |
| 111 | +``` |
| 112 | + |
| 113 | +## What Gets Checked |
| 114 | + |
| 115 | +- **CREATE TABLE** statements are validated by the event trigger |
| 116 | +- **ALTER TABLE** is deliberately skipped — you can't reorder existing columns, so warning would be noise |
| 117 | +- **Temp tables** and **system schemas** (`pg_catalog`, `information_schema`) are skipped |
| 118 | +- Tables in the `exclusions` list are skipped |
| 119 | + |
| 120 | +## Requirements |
| 121 | + |
| 122 | +- PostgreSQL 9.5+ (requires `pg_event_trigger_ddl_commands()`) |
| 123 | +- Superuser privileges to create the event trigger |
| 124 | + |
| 125 | +## License |
| 126 | + |
| 127 | +MIT |
0 commit comments