Internal TOAST dan Penyimpanan Nilai Besar

PostgreSQL stores rows on fixed-size 8 KB pages. A single row cannot span multiple pages, so a wide value (a long text, big jsonb, or bytea) would never fit.

TOAST (The Oversized-Attribute Storage Technique) solves this by compressing oversized columns and, if still too large, slicing them into chunks stored in a separate side table.

• Keeps the main heap row small and cache-friendly.
• Lets a logical value far exceed 8 KB (up to ~1 GB).
• Happens automatically and transparently to your queries.

TOAST kicks in when a row's total size would exceed TOAST_TUPLE_THRESHOLD, which is 2 KB (one quarter of the 8 KB page) by default.

When that limit is crossed, PostgreSQL compresses and/or moves the largest toastable attributes out of line until the row fits under TOAST_TUPLE_TARGET (also ~2 KB).

Only columns of variable-length types (text, varchar, jsonb, bytea, arrays, etc.) are toastable. Fixed-width types like integer or timestamptz are never toasted.

Every table with at least one toastable column gets an associated TOAST table named pg_toast.pg_toast_<oid>. You can locate it from the catalog.

The reltoastrelid column links a heap relation to its TOAST relation; a value of 0 means no TOAST table was created.

Each column has a storage strategy controlling whether it can be compressed and/or moved out of line:

• PLAIN — no compression, no out-of-line; only valid for non-toastable types.
• EXTENDED — allow both compression and out-of-line storage (default for most varlena types).
• EXTERNAL — allow out-of-line but no compression (faster substring access).
• MAIN — allow compression but keep in the main table unless absolutely necessary.

Query pg_attribute.attstorage to see each column's strategy. The codes map to: p=PLAIN, e=EXTERNAL, m=MAIN, x=EXTENDED.

Use ALTER TABLE ... SET STORAGE to override the default. A common optimization: if you frequently read random substrings of a large bytea (e.g. range reads on a blob), switch to EXTERNAL so values are stored uncompressed and can be partially fetched without decompressing the whole thing.

The new strategy applies only to rows written after the change; existing data is not rewritten until updated.

PostgreSQL compresses TOAST values before considering out-of-line storage. Two algorithms are available:

• pglz — the historic built-in, decent ratio, slower.
• lz4 — available since PostgreSQL 14, much faster compress/decompress with a slightly lower ratio. Requires the server to be built with lz4 support.

The cluster-wide default is set by default_toast_compression.

From PostgreSQL 14 onward you can set the compression method on individual columns with SET COMPRESSION. This is independent of the storage strategy.

Choose lz4 for hot, large columns where CPU during read/write matters; keep pglz or use EXTERNAL when ratio or substring access dominates.

The reloption toast_tuple_target controls how aggressively PostgreSQL pushes attributes out of line: it sets the size the main tuple is shrunk toward (valid range 128 bytes to ~8160).

Lowering it makes more values go to TOAST sooner, keeping the main heap dense and improving scan performance when the big columns are rarely read. Raising it keeps more data inline.

Use the size functions to separate main-table bytes from TOAST bytes. This tells you whether your bloat lives in the heap or in oversized values.

• pg_table_size — heap + TOAST + their indexes' TOAST, excluding regular indexes.
• pg_relation_size(rel, 'main') — just the main fork.

TOAST is invisible until it isn't. Key effects to remember:

• Reading a toasted column triggers extra index lookups into the TOAST table and possible decompression — avoid SELECT * when you only need small columns.
• Values stay TOASTed across UPDATEs that don't touch them, so unrelated updates are cheap.
• EXTERNAL enables efficient substr() / range reads on large uncompressed blobs.
• Switching pglz to lz4 can cut write CPU dramatically on insert-heavy large-value workloads.

You have a table whose large bytea column is read mostly via substr() on small byte ranges, and these reads are slow because each access decompresses the entire value. Which single change best fixes this?

You learned how PostgreSQL handles oversized values:

• TOAST triggers when a row would exceed the ~2 KB threshold, compressing then moving large varlena columns into pg_toast.* tables.
• Four strategies — PLAIN, MAIN, EXTENDED (default), EXTERNAL — control compression and out-of-line placement, set via ALTER TABLE ... SET STORAGE.
• lz4 (PG14+) offers faster compression than pglz; choose per column with SET COMPRESSION or cluster-wide via default_toast_compression.
• toast_tuple_target tunes how eagerly values leave the heap; size functions reveal how much storage lives in TOAST.
• Pick EXTERNAL for substring/range reads, lz4 for write-heavy large values, and avoid SELECT * to skip needless detoasting.