Postgresql on Murat Useinov

Testing PostgreSQL 19 beta without guessing

Wed, 08 Jul 2026 21:45:00 +0300

A production-like dump, a separate PostgreSQL 19 beta instance, one evening plus machine time. That is the whole method, and it is the only thing a database beta is for: finding out where the new version changes the behavior of your application while the change is still a bug report and not your incident.

Restore the dump first. Data size matters. A plan that is fine on a thousand rows goes bad on fifty million.

PostgreSQL 18, the upgrade notes

Fri, 26 Sep 2025 22:45:00 +0300

PostgreSQL 18 went GA yesterday. I ran the beta in July. Now it is not a preview, it is a ticket, and these are the notes from planning it.

The quiet feature that matters most for the upgrade itself: pg_upgrade now carries planner statistics over. Before, the minutes or hours after switching versions were the scary part. Fresh cluster, empty statistics, the planner guessing, production limping until ANALYZE finished. That cliff is mostly gone. This alone changes how a major upgrade feels at 2 a.m.

A weekend with PostgreSQL 18 beta

Wed, 02 Jul 2025 18:50:00 +0300

A staging copy of one database, restored over the weekend onto PostgreSQL 18 beta. GA is expected in autumn. Two things I wanted to touch: asynchronous I/O and the built-in uuidv7().

Start with uuidv7, because it is about a mistake many of us already made, me included. Random UUIDv4 primary keys scatter inserts across the whole index. Every insert lands on a random page, the working set is the entire index, buffers churn. UUIDv7 is time-ordered: new rows go to the same few pages, like a sequence, but still globally unique.

PostgreSQL 17 and why VACUUM is an API concern

Fri, 27 Sep 2024 15:20:00 +0300

A reporting transaction open for hours, and n_dead_tup on the hot table climbing all day while autovacuum ran non-stop and changed nothing.

That was one project. PostgreSQL 17 came out yesterday, and the headline work is exactly this plumbing: vacuum got a new memory structure for tracking dead tuples, far less memory and no longer capped the old way, plus better throughput for high-concurrency writes and bulk loads. Maintenance internals. Which is why backend people should care.

PostgreSQL 16, read through one workload

Sat, 16 Sep 2023 20:25:00 +0300

A nightly import, a few hundred million rows through COPY, then reporting queries in the morning. That is the only workload I care about, so that is how I read the PostgreSQL 16 release notes: with one pipeline in mind and everything else skipped.

Three items survived the filter.

COPY got faster, noticeably when several loads run at once. Our import is exactly this shape, parallel COPY streams into partitioned tables. I will take the free speed. But the old lesson stays: if your import still does row-by-row INSERT from PHP, no Postgres release will save you. Batch into COPY first, then talk about versions.

UUID primary keys and the B-tree bill

Thu, 03 Aug 2023 21:35:00 +0300

The same table, the same rows, about 1.5 times the disk. That was the measurement after someone made UUIDv4 the primary key in InnoDB, and that is before counting secondary indexes.

Every new API wants UUIDs. Distributed generation, no coordination, nothing leaks about row counts. All true. Then the database quietly starts paying.

InnoDB clusters the table by primary key. Auto-increment inserts always land in the rightmost page, warm and predictable. Random UUIDs land anywhere in the tree: page splits everywhere, half-empty pages, and the set of hot pages becomes the whole index. Secondary indexes store the primary key in every entry, so each of them inherits the full 16 bytes too. In PostgreSQL the heap does not cluster, so the pain is smaller, but index bloat and lost locality are still there.

MERGE arrives in PostgreSQL 15

Thu, 27 Oct 2022 20:20:00 +0300

A supplier price list, loaded into staging_prices, then reconciled against prices. Since PostgreSQL 15 came out two weeks ago, that is one statement:

MERGE INTO prices p
USING staging_prices s ON p.sku = s.sku
WHEN MATCHED AND s.price IS NULL THEN
    DELETE
WHEN MATCHED AND p.price <> s.price THEN
    UPDATE SET price = s.price, updated_at = now()
WHEN NOT MATCHED THEN
    INSERT (sku, price) VALUES (s.sku, s.price);

Before 15 this was three statements in a transaction, or a stored procedure, or a loop in PHP. Now it is one statement that says what it does. Update changed rows, delete withdrawn ones, insert new ones. People coming from Oracle and SQL Server waited a decade for this. The conditional WHEN MATCHED AND ... branches are the real value. ON CONFLICT cannot express “delete when the source says so” at all.

JSONB or a normal table

Sat, 16 Jul 2022 10:05:00 +0300

Audit snapshots, product attributes, and a document with two hot keys. Three cases from one Postgres project, one question for each: column or JSONB.

I answer it by looking at queries, not at data.

Audit snapshots. When an order changes, we store the whole previous state. Nobody queries inside it. It is read as a blob, whole, rarely, by a human during an incident. Perfect JSONB. Normalizing it would mean a dozen tables for data nobody joins.

Backfill in batches, not in one UPDATE

Thu, 17 Mar 2022 22:40:00 +0300

Sixty million rows in orders, a new nullable column total_cents from expand-and-contract, and a Friday afternoon. The question is how to fill it.

The naive answer is one statement. UPDATE orders SET total_cents = ROUND(total * 100). On a test database it works. On production it holds row locks on the whole table for the whole run, writes one giant chunk of WAL, and the replica falls minutes behind. On MySQL the binlog gets the same present. One project I worked on learned this on that Friday. Since then I backfill only in batches.

PostgreSQL 14 and the price of a connection

Thu, 30 Sep 2021 10:05:00 +0300

A few hundred connections, mostly idle, and the whole server gets slower. That was Postgres for as long as I remember it: a process per connection, and past some count the snapshot bookkeeping starts to eat everything.

PostgreSQL 14 is released today. I had the RC on a test box for a week, so a few notes from a backend seat. The headline for me is that this internal work got a serious rewrite. On our connection-heavy profile, many workers, short queries, lots of idle time, the difference is visible without squinting. Idle connections finally cost close to what people always assumed they cost.

The outbox table

Tue, 23 Mar 2021 15:30:00 +0300

About once a month an order existed and nobody heard about it. The order service saved the order, then published order.created to RabbitMQ, and once a month the process died between the two. Rare enough to be mysterious. Frequent enough to ruin a weekend.

Two operations, two systems, no common transaction. Whatever order you do them in, you lose. Commit first, then publish: the order exists, the event never leaves. Publish first, then commit: the commit fails, consumers are already busy with an order that does not exist. We had the first variant in production.

A deadlock is the database doing its job

Sat, 02 May 2020 20:30:00 +0300

Error 1213 in the logs, and the chat lights up: the database is broken. It is not. Two transactions locked rows in opposite order, A waits for B, B waits for A, and InnoDB did the only sane thing: picked a victim and killed it. The cycle is gone. This is a feature.

The classic shape is a money transfer. One request moves funds from account 1 to account 2, another from 2 to 1, both update the first account and then the second. Opposite order, instant cycle under load. Tests never show it, because tests do not run two of these in the same millisecond.

PostgreSQL 12 inlines your CTE

Fri, 18 Oct 2019 21:45:00 +0300

grep -rn "WITH " src/ was the first thing I ran after upgrading to PostgreSQL 12 this month. I was looking for a trick that had stopped working.

For years WITH was an optimization fence. The planner materialized the CTE first and only then ran the outer query. Everyone used this both ways. As a bug: you wrap a subquery into a CTE for readability, the planner stops pushing conditions inside, a fast query becomes a scan of half a table. As a feature: you write a CTE on purpose, to pin evaluation order and stop the planner from being creative. Half of the CTE advice on the internet is really advice about the fence.

Transactional DDL is not portable

Mon, 18 Mar 2019 22:45:00 +0300

Statement three of a migration fails. Column name typo. On PostgreSQL the first two ALTERs roll back with it, the schema returns to the exact state before the migration, you fix the typo and run again. On MySQL the first two ALTERs are already permanent.

Same up() method, same php artisan migrate, same green output. The migration tool gives one abstraction over two very different databases. The abstraction covers syntax. It does not cover what happens on failure.

PostgreSQL 11: partitions grow up, JIT arrives

Fri, 26 Oct 2018 19:10:00 +0300

jit = on, restart, run the API test suite. Same numbers as before. That was my first evening with PostgreSQL 11, released last week, and it was the correct result.

JIT is the loud feature of this release and the misunderstood one. Postgres can now compile expression evaluation into machine code through LLVM. People read “compilation” and expect their endpoints to get faster. They will not. A primary key lookup takes a fraction of a millisecond. There is nothing in it worth compiling, and the compilation itself costs more than the whole query. JIT is for the other kind of query: an aggregate chewing through millions of rows, where the same expression runs so many times that generating machine code for it pays back.

Covering index and the price of SELECT *

Sun, 15 Jul 2018 12:35:00 +0300

Using index in the Extra column of EXPLAIN. That is the cheapest read MySQL can do, and you often get it almost for free.

SELECT user_id, created_at
FROM orders
WHERE user_id = 42
ORDER BY created_at DESC
LIMIT 20;

With an index on (user_id, created_at) everything the query needs is in the index leaves. The table is never touched. A composite index finds rows fast, a covering index answers the query on its own. Postgres calls this Index Only Scan, with one condition: the visibility map must be fresh, so a table that vacuum never visits quietly falls back to heap fetches.

PostgreSQL 10: partitions I can finally use

Thu, 26 Oct 2017 19:20:00 +0300

DROP TABLE events_2017_10. That is how you delete October now, and it is the reason I am reading PostgreSQL 10 release notes on a Thursday evening instead of waiting a year like usual.

Partitioning existed before, through inheritance, CHECK constraints and an insert trigger you wrote yourself and hoped was right. Now the database owns it:

CREATE TABLE events (
    id bigserial NOT NULL,
    created_at timestamptz NOT NULL,
    payload jsonb
) PARTITION BY RANGE (created_at);

CREATE TABLE events_2017_10 PARTITION OF events
    FOR VALUES FROM ('2017-10-01') TO ('2017-11-01');

The win is the data lifecycle. An events table grows forever, and deleting a year of history with DELETE is a night of I/O plus a bloated table in the morning. With partitions, retiring a month is one DROP. Instant. And a query that filters by created_at visits only the partitions in range, the planner skips the rest.

Keyset pagination instead of OFFSET

Mon, 19 Jun 2017 10:20:00 +0300

LIMIT 50 OFFSET 500000. Page 10001 of an events table, from a paginator someone wrote in an afternoon. Postgres has no shortcut to row 500000. It walks the index through half a million entries, fetches them, throws them away, and returns fifty. Page one is fast. Page ten thousand is slow, and every page after it is slower. Run EXPLAIN ANALYZE on both: the plans are identical, the numbers are not.

Parallel query in PostgreSQL 9.6

Fri, 30 Sep 2016 22:10:00 +0300

38 seconds to 11. Same query, same forty million rows, no index added, no SQL changed. The only difference is PostgreSQL 9.6, released yesterday, and one setting.

The query is a typical report: count and sum over an events table, grouped by day, three months of data. On 9.5 the plan is one process grinding through the table. On 9.6 with parallelism enabled:

Finalize HashAggregate
  ->  Gather
        Workers Planned: 4
        ->  Partial HashAggregate
              ->  Parallel Seq Scan on events

Four workers scan their own chunks, aggregate partially, the leader merges. Just more hands. For years we optimized SQL as if the database had exactly one worker per query. That assumption expired yesterday.

Upsert in PostgreSQL 9.5

Tue, 12 Jan 2016 22:15:00 +0300

$row = $db->fetchOne('SELECT id FROM counters WHERE name = ?', array($name));
if ($row) {
    $db->execute('UPDATE counters SET value = value + 1 WHERE name = ?', array($name));
} else {
    $db->execute('INSERT INTO counters (name, value) VALUES (?, 1)', array($name));
}

Works on the laptop. In production two requests arrive in the same millisecond. Both SELECT, both see nothing, both INSERT. One dies with a duplicate key error. Or worse, there is no unique constraint, and now you have two rows and a bug report you cannot reproduce. The window between SELECT and INSERT is tiny, so it fires once a week, always for someone else.