Snapshots under the hood
Chapter 2 treated the snapshot as a black box: "a frozen view of committed data". Here's the box, opened. A snapshot is three numbers, and every visibility decision (every row you do or don't see) is arithmetic on them plus the xmin/xmax stamps from the previous lesson.
| component | meaning |
|---|---|
xmin | the oldest xid still running when the snapshot was taken, everything below is settled (committed or rolled back) |
xmax | one past the highest completed xid, everything at or above hadn't finished, so it's invisible |
xip | the xids between the two that were still in progress at snapshot time: invisible, forever, in this snapshot |
A row version is visible if its xmin is committed, below xmax, and not in xip, and its xmax (if set) fails that same test. Notice what's absent from the rule: the current time. Whether the creator commits a nanosecond or an hour after the snapshot was taken changes nothing.
The three numbers, live
Transaction ids are handed out lazily — a transaction that only reads never gets one.
A> BEGIN;
BEGIN
A> SELECT balance FROM accounts WHERE id = 1;
balance
---------
100
(1 row)
A> SELECT pg_current_xact_id_if_assigned() AS xid;
xid
-----
(1 row)
A> UPDATE accounts SET balance = 150 WHERE id = 1;
UPDATE 1
A> SELECT pg_current_xact_id_if_assigned() AS xid;
xid
------
1001
(1 row)B grabs the next xid and commits immediately. A is still in progress.
B> UPDATE accounts SET balance = 200 WHERE id = 2 RETURNING xmin, balance;
xmin | balance
------+---------
1002 | 200
(1 row)C opens a transaction and inspects its own snapshot: the three numbers that decide all visibility.
C> BEGIN ISOLATION LEVEL REPEATABLE READ;
BEGIN
C> SELECT pg_snapshot_xmin(pg_current_snapshot()) AS xmin,
pg_snapshot_xmax(pg_current_snapshot()) AS xmax;
xmin | xmax
------+------
1001 | 1003
(1 row)xip = the xids that were in progress at snapshot time. A is on the list — so A is invisible even after it commits.
C> SELECT pg_snapshot_xip(pg_current_snapshot()) AS xid;
xid
------
1001
(1 row)The arithmetic in action: B (committed, below xmax) is visible; A (in xip) is not.
C> SELECT id, owner, balance FROM accounts ORDER BY id;
id | owner | balance
----+-------+---------
1 | alice | 100
2 | bob | 200
(2 rows)
A> COMMIT;
COMMITA has now committed — but C's snapshot already decided: still invisible.
C> SELECT id, owner, balance FROM accounts ORDER BY id;
id | owner | balance
----+-------+---------
1 | alice | 100
2 | bob | 200
(2 rows)
C> COMMIT;
COMMITOnly a NEW snapshot changes the verdict.
C> SELECT id, owner, balance FROM accounts ORDER BY id;
id | owner | balance
----+-------+---------
1 | alice | 150
2 | bob | 200
(2 rows)Verified against PostgreSQL 18.4 · Run it yourself · Scenario source
Three beats reward a second look. The first is that xids are handed out lazily: A's transaction had no xid until its first write, because assignment "happens when a transaction first writes to the database". Read-only transactions never consume one, which is why monitoring columns like backend_xid are so often null for busy read-only sessions, and why the workloads that burn through xids are the write-heavy ones.
The second: xmax is one past the highest completed xid, not the next id to be assigned. A got its xid before B, yet sits at xmin, below xmax, alive in xip: the three numbers bracket exactly the region where "did this commit?" is still ambiguous.
The third is the punchline: commit doesn't matter, because the snapshot already decided. A commits mid-scenario and C's verdict doesn't budge. That reframes the whole of chapter 2: the isolation levels are nothing but retention policies for this one object. Read Committed takes a fresh snapshot per statement; Repeatable Read keeps one for the whole transaction. Same arithmetic, different lifetime.
That's the lever, then: inspect your own snapshot with pg_current_snapshot(), and remember that visibility is pure arithmetic on xids: no clocks, no "latest committed value" lookup. It's why MVCC reads are cheap, and also why old snapshots are expensive: someone has to keep every version they might still need, which is the whole problem the long-transactions lesson picks apart next.