PostgreSQL Performance & Query Optimization · 课时

在线将超大表迁移到分区表

在极少锁定且不丢失数据的情况下，将现有单体表转换为分区表。

在线将超大表迁移到分区表 是 CoddyKit 上的免费 PostgreSQL Performance & Query Optimization 课时。这是第 4 节课，共 4 节。你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。这是 PostgreSQL Performance & Query Optimization 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 PostgreSQL Performance & Query Optimization 课程共包含 4 节课。

本课时的部分内容尚未翻译，以英文显示。

Why Migrate a Huge Table?

Imagine an events table holding 800 million rows of append-only log data. Every query scans a monstrous index, VACUUM runs for hours, and dropping old data with DELETE bloats the table.

Partitioning splits one logical table into many physical child tables based on a key (for example, created_at by month). Benefits include:

Partition pruning — the planner skips irrelevant partitions entirely.
Instant data retention — DROP or DETACH a whole month in milliseconds, no row-by-row delete.
Cheaper maintenance — VACUUM and reindex run per partition.

The challenge: doing this on a live, write-heavy table without long locks or data loss.

The Naive Approach and Its Trap

PostgreSQL cannot turn an existing plain table into a partitioned one with a single ALTER TABLE. A partitioned parent is a different kind of object created with PARTITION BY.

The tempting one-shot plan is: create the partitioned table, then move all rows in a single transaction.

This blocks the table with heavy locks for the entire copy and holds a giant transaction open. On 800M rows that means hours of downtime and enormous WAL. We need an online strategy instead.

-- This single INSERT...SELECT locks and runs for hours.
-- Holds one transaction open across the whole 800M-row copy.
INSERT INTO events_partitioned
SELECT * FROM events_old;  -- DON'T do this on a live huge table

Strategy Overview: Shadow Table + Backfill + Swap

The proven online recipe has four phases:

Create a new partitioned shadow table with matching columns and a partition key.
Dual-write — keep new rows flowing into both old and new tables via a trigger (or write to the new one once it exists).
Backfill historical rows in small, committed batches so locks stay short.
Swap names inside one short transaction, then drop the old table.

Each phase is individually safe and resumable. No single long-running lock, no lost writes.

Step 1: Create the Partitioned Shadow

Create a new parent declared with PARTITION BY RANGE on the chosen key. The partition key column must be part of the primary key in declarative partitioning.

We define monthly range partitions. Note that the parent itself stores no rows; each child owns a slice.

CREATE TABLE events_new (
    id          bigint        GENERATED ALWAYS AS IDENTITY,
    user_id     bigint        NOT NULL,
    event_type  text          NOT NULL,
    payload     jsonb,
    created_at  timestamptz   NOT NULL DEFAULT now(),
    PRIMARY KEY (id, created_at)   -- partition key must be in PK
) PARTITION BY RANGE (created_at);

CREATE TABLE events_new_2024_01 PARTITION OF events_new
    FOR VALUES FROM ('2024-01-01') TO ('2024-02-01');

CREATE TABLE events_new_2024_02 PARTITION OF events_new
    FOR VALUES FROM ('2024-02-01') TO ('2024-03-01');

Step 2: A Default Partition as a Safety Net

If a row's key falls outside every defined range, the INSERT fails. While migrating you may not have created every month yet, so add a default partition to catch stragglers.

Watch out: attaching a new partition later requires PostgreSQL to scan the default to prove no conflicting rows exist. Keep the default empty in steady state by pre-creating the months you actually need.

CREATE TABLE events_new_default
    PARTITION OF events_new DEFAULT;

-- Later, when you add a real partition, PostgreSQL scans
-- the default for conflicting rows before attaching.
CREATE TABLE events_new_2024_03 PARTITION OF events_new
    FOR VALUES FROM ('2024-03-01') TO ('2024-04-01');

Step 3: Keep New Writes in Sync

While we backfill the past, the application keeps inserting. We must not lose those live rows. A robust pattern is a trigger on the old table that mirrors every write into the new partitioned table.

Once backfill is done and swap is near, the trigger guarantees both tables stay identical for fresh data.

CREATE OR REPLACE FUNCTION mirror_to_new()
RETURNS trigger AS $$
BEGIN
    INSERT INTO events_new (user_id, event_type, payload, created_at)
    VALUES (NEW.user_id, NEW.event_type, NEW.payload, NEW.created_at);
    RETURN NEW;
END;
$$ LANGUAGE plpgsql;

CREATE TRIGGER trg_mirror_events
    AFTER INSERT ON events_old
    FOR EACH ROW EXECUTE FUNCTION mirror_to_new();

Step 4: Backfill in Small Batches

Now copy historical rows in bounded, committed chunks. Each batch is its own transaction, so locks release immediately and you can pause or resume anytime.

Drive the loop by the primary key or a timestamp window. Use ON CONFLICT DO NOTHING so rows already mirrored by the trigger don't cause duplicates.

-- Run repeatedly (from a script) until 0 rows are moved.
INSERT INTO events_new (id, user_id, event_type, payload, created_at)
SELECT id, user_id, event_type, payload, created_at
FROM   events_old
WHERE  id > :last_id
ORDER  BY id
LIMIT  10000
ON CONFLICT (id, created_at) DO NOTHING;

-- Capture MAX(id) of this batch as the next :last_id, then COMMIT.

Why Batching Beats One Big Copy

Small batches matter for concrete reasons:

Lock duration — each batch holds row locks for milliseconds, not hours.
WAL and bloat — committed batches let VACUUM and checkpoints keep up; one giant transaction balloons WAL.
Replication lag — replicas apply small chunks smoothly instead of stalling on a huge transaction.
Resumability — a crash mid-migration loses only the current batch.

Throttle with a short pg_sleep between batches if you see I/O pressure or replica lag.

-- Optional throttle between batches to ease I/O / replica lag.
SELECT pg_sleep(0.2);

Step 5: Reconcile and Add Indexes

Before the swap, verify both tables agree and build the indexes the new table needs.

On a partitioned table, creating an index on the parent automatically creates matching indexes on every partition. Use CREATE INDEX (it cascades) and build it once backfill is complete to avoid slowing the copy.

-- Sanity check: counts should match (allow for in-flight writes).
SELECT (SELECT count(*) FROM events_old)  AS old_count,
       (SELECT count(*) FROM events_new)  AS new_count;

-- Cascades to all current and future partitions.
CREATE INDEX idx_events_new_user_id
    ON events_new (user_id);

CREATE INDEX idx_events_new_created_at
    ON events_new (created_at);

Step 6: The Atomic Swap

The cutover is a single short transaction that renames tables. Because RENAME only changes catalog entries, it takes an ACCESS EXCLUSIVE lock for a tiny moment.

Drop the mirror trigger first (the new table is about to become the real one), do a final catch-up batch, then rename. Keep this transaction minimal.

BEGIN;

DROP TRIGGER trg_mirror_events ON events_old;

-- Final tiny catch-up for any rows written since last batch.
INSERT INTO events_new (id, user_id, event_type, payload, created_at)
SELECT id, user_id, event_type, payload, created_at
FROM   events_old
ON CONFLICT (id, created_at) DO NOTHING;

ALTER TABLE events_old RENAME TO events_retired;
ALTER TABLE events_new RENAME TO events;

COMMIT;

Step 7: Verify, Then Clean Up

After the swap, confirm the live table is partitioned and pruning works. EXPLAIN a date-bounded query: only the relevant partitions should appear.

Keep events_retired around for a short safety window, then drop it to reclaim space. Going forward, automate creating next month's partition ahead of time.

-- Should touch only Jan/Feb partitions, not the whole table.
EXPLAIN (COSTS OFF)
SELECT count(*) FROM events
WHERE created_at >= '2024-01-10'
  AND created_at <  '2024-02-05';

-- After a safe verification window:
DROP TABLE events_retired;

Quick Check: Choosing the Cutover Strategy

You are migrating a 1-billion-row, write-heavy table to monthly range partitions with the smallest possible disruption. Which final cutover approach is correct?

Recap: Online Partition Migration

You converted a monolithic table into partitions with no downtime by:

Creating a partitioned shadow table with the partition key in the primary key.
Adding a default partition as a safety net and pre-creating needed ranges.
Installing a mirror trigger so live writes hit both tables.
Backfilling in small committed batches with ON CONFLICT DO NOTHING for short locks and resumability.
Building indexes on the parent (they cascade to partitions) and reconciling counts.
Performing an atomic RENAME swap after a final catch-up, then dropping the retired table.

The guiding principle: never hold one long lock or one giant transaction — split the work so the live system keeps serving traffic throughout.

常见问题解答

「在线将超大表迁移到分区表」课时是免费的吗？

是的 — 「在线将超大表迁移到分区表」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 PostgreSQL Performance & Query Optimization 课程的其余内容，请升级到 CoddyKit PRO。 PostgreSQL Performance & Query Optimization 课程共包含 4 节课。

「在线将超大表迁移到分区表」这节课中我会学到什么？

在极少锁定且不丢失数据的情况下，将现有单体表转换为分区表。你通过在浏览器中直接运行的动手代码来练习 PostgreSQL Performance & Query Optimization，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 PostgreSQL Performance & Query Optimization 需要有经验吗？

无需任何先前经验。CoddyKit 上的 PostgreSQL Performance & Query Optimization 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。这是第 4 节课，共 4 节。

「在线将超大表迁移到分区表」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 PostgreSQL Performance & Query Optimization 课中编写并运行代码吗？

能。每节 PostgreSQL Performance & Query Optimization 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

← 返回 PostgreSQL Performance & Query Optimization