การทำ Upsert ขนาดใหญ่ด้วย ON CONFLICT

An upsert inserts a row, but if it would collide with an existing key, it updates the existing row instead. PostgreSQL spells this INSERT ... ON CONFLICT.

For small workloads it is trivial. For ETL-sized batches (tens of thousands to millions of rows) the naive approach causes three problems:

• Lock contention — concurrent writers fighting over the same rows or index pages.
• Table bloat — every UPDATE writes a new row version (dead tuple) that VACUUM must later reclaim.
• WAL and round-trip overhead — row-by-row upserts multiply network and transaction cost.

This lesson builds a merge that is both correct and throughput-friendly.

ON CONFLICT requires a conflict target: the column(s) or constraint that define a duplicate. PostgreSQL needs a unique or exclusion constraint on that target to arbitrate.

The two actions are DO NOTHING (skip the colliding row) and DO UPDATE (merge new values in).

Inside DO UPDATE, the incoming row is exposed through the special EXCLUDED pseudo-table.

The single most important throughput rule: batch your rows into one statement. A multi-row VALUES list (or a feeding SELECT) is parsed, planned, and committed once instead of N times.

This collapses N network round-trips and N transaction commits into one, often a 10-100x speedup over row-by-row upserts.

For real ETL, load raw data into an unlogged staging table first (often via COPY, the fastest bulk path), then merge from staging into the target with one INSERT ... SELECT ... ON CONFLICT.

Benefits:

• COPY avoids per-row INSERT overhead.
• An UNLOGGED staging table skips WAL for the load phase.
• You can dedupe and transform in the SELECT before merging.

A subtle but fatal error: a single INSERT statement cannot update the same target row twice. If your batch contains two rows with the same conflict key, PostgreSQL raises:

ERROR: ON CONFLICT DO UPDATE command cannot affect row a second time

Fix it by collapsing duplicates in the source before merging. DISTINCT ON keeps one row per key — typically the newest.

Every DO UPDATE writes a new row version, even when the new values are identical to the old ones. Those dead tuples bloat the table and create extra work for VACUUM.

Add a WHERE clause to the DO UPDATE so it fires only when something actually changed. Use IS DISTINCT FROM so NULLs compare correctly.

When several ETL workers run concurrently, deadlocks appear if they touch the same keys in different orders. Worker 1 locks key X then Y; worker 2 locks Y then X — both block forever until PostgreSQL kills one.

Defenses:

• Sort each batch by the conflict key so all workers acquire locks in the same order.
• Partition work by key range so no two workers share keys.
• Keep transactions short — long-held row locks magnify contention.

One enormous transaction merging millions of rows holds locks for a long time, balloons WAL, and blocks autovacuum from cleaning up. Break the work into chunks (for example 10k-50k rows) and commit between them.

Smaller transactions release locks sooner, let autovacuum keep pace, and make retries cheap after a failure. The trade-off is slightly more commit overhead — tune the chunk size against your hardware.

The conflict target must match an actual unique/primary-key or exclusion constraint. You can target by column list ON CONFLICT (sku) or by constraint name ON CONFLICT ON CONSTRAINT products_sku_key.

For partial unique indexes, repeat the index predicate so PostgreSQL can pick the right one:

PostgreSQL 15+ adds the SQL-standard MERGE, which can INSERT, UPDATE, and DELETE in one pass. For high-throughput upserts, INSERT ... ON CONFLICT is usually still preferred:

• ON CONFLICT is atomic against concurrent inserts — it handles a race where another transaction inserts the same key, retrying internally.
• Classic MERGE can raise a unique-violation under heavy concurrency because it does not have that built-in conflict arbitration.

Use MERGE when you need DELETE branches or complex conditional logic; use ON CONFLICT for plain, concurrency-safe upserts.

Even an optimized merge produces dead tuples on the updated rows. Keep performance steady with maintenance discipline:

• Ensure autovacuum keeps up; for hot ETL tables lower autovacuum_vacuum_scale_factor so it triggers more often.
• After a massive one-off backfill, run VACUUM (ANALYZE) to reclaim space and refresh planner statistics.
• Every extra index on the target slows the merge (each insert/update maintains all of them) — keep only the indexes you truly need.

Test your understanding of safe, high-throughput merges.

Efficient upserts at scale come from a few combined habits:

• Batch rows into one statement; for ETL, stage with COPY then INSERT ... SELECT ... ON CONFLICT.
• Deduplicate the batch (DISTINCT ON) so no key appears twice.
• Skip no-op updates with a WHERE ... IS DISTINCT FROM clause to curb dead tuples and bloat.
• Order by the conflict key and partition work to avoid deadlocks; chunk huge merges into short transactions.
• Prefer ON CONFLICT for concurrency-safe upserts; keep autovacuum healthy and indexes lean.

Together these turn a fragile row-by-row merge into a fast, low-contention ETL load.