This is an enterprise-only feature. Request a 30-day trial license to try it out.
READ COMMITTED
is one of two transaction isolation levels supported on CockroachDB. By default, CockroachDB uses the SERIALIZABLE
isolation level, which is the strongest ANSI transaction isolation level.
READ COMMITTED
isolation is appropriate in the following scenarios:
Your application needs to maintain a high workload concurrency with minimal transaction retries, and it can tolerate potential concurrency anomalies. Predictable query performance at high concurrency is more valuable than guaranteed transaction serializability.
You are migrating an application to CockroachDB that was built at a
READ COMMITTED
isolation level on the source database, and it is not feasible to modify your application to useSERIALIZABLE
isolation.
Whereas SERIALIZABLE
isolation guarantees data correctness by placing transactions into a serializable ordering, READ COMMITTED
isolation permits some concurrency anomalies in exchange for minimizing transaction aborts, retries, and blocking. Compared to SERIALIZABLE
transactions, READ COMMITTED
transactions do not return serialization errors that require client-side handling. See READ COMMITTED
transaction behavior.
If your workload is already running well under SERIALIZABLE
isolation, Cockroach Labs does not recommend changing to READ COMMITTED
isolation unless there is a specific need.
READ COMMITTED
on CockroachDB provides stronger isolation than READ COMMITTED
on PostgreSQL. On CockroachDB, READ COMMITTED
prevents anomalies within single statements. For complete details on how READ COMMITTED
is implemented on CockroachDB, see the Read Committed RFC.
Enable READ COMMITTED
isolation
By default, the sql.txn.read_committed_isolation.enabled
cluster setting is true
, enabling READ COMMITTED
transactions. If the cluster setting is false
, READ COMMITTED
transactions will run as SERIALIZABLE
.
To check whether any transactions are being upgraded to SERIALIZABLE
, see the Upgrades of SQL Transaction Isolation Level graph in the DB Console.
Set the default isolation level to READ COMMITTED
To set all future transactions to run at READ COMMITTED
isolation, use one of the following options:
The
SET SESSION CHARACTERISTICS
statement, which applies to the current session:SET SESSION CHARACTERISTICS AS TRANSACTION ISOLATION LEVEL READ COMMITTED;
The
default_transaction_isolation
session variable:At the session level:
SET default_transaction_isolation = 'read committed';
At the database level:
ALTER DATABASE db SET default_transaction_isolation = 'read committed';
At the role level:
ALTER ROLE foo SET default_transaction_isolation = 'read committed';
The
default_transaction_isolation
session variable as a connection parameter withcockroach sql
:cockroach sql -–url='postgresql://{username}@{host}:{port}/{database}?options=-c default_transaction_isolation=read%20committed'
To view the default isolation level of the session:
SHOW default_transaction_isolation;
default_transaction_isolation
-----------------------------------------------------
read committed
Set the current transaction to READ COMMITTED
To begin a transaction as a READ COMMITTED
transaction, use one of the following options:
The
BEGIN TRANSACTION ISOLATION LEVEL
statement:BEGIN TRANSACTION ISOLATION LEVEL READ COMMITTED;
The
SET TRANSACTION ISOLATION LEVEL
statement, at the beginning of the transaction:BEGIN; SET TRANSACTION ISOLATION LEVEL READ COMMITTED;
The
transaction_isolation
session variable, at the beginning of the transaction:BEGIN; SET transaction_isolation = 'read committed';
To view the isolation level of a transaction, run SHOW
within the open transaction:
SHOW transaction_isolation;
transaction_isolation
-------------------------
read committed
Starting a transaction as READ COMMITTED
does not affect the default isolation level, which can be different.
READ COMMITTED
transaction behavior
READ COMMITTED
and SERIALIZABLE
transactions both serve globally consistent ("non-stale") reads and commit atomically. READ COMMITTED
transactions have the following differences:
Writes in concurrent
READ COMMITTED
transactions can interleave without aborting transactions, and a write can never block a non-locking read of the same row. This is becauseREAD COMMITTED
transactions are not required to be placed into a serializable ordering.Whereas statements in
SERIALIZABLE
transactions see data that committed before the transaction began, statements inREAD COMMITTED
transactions see data that committed before each statement began. If rows are being updated by concurrent writes, reads in aREAD COMMITTED
transaction can return different results.Note:For details on how this is implemented, see Read snapshots.
Due to the preceding behaviors,
READ COMMITTED
transactions permit some types of concurrency anomalies that are prevented inSERIALIZABLE
transactions. For details and examples, see Concurrency anomalies.You can mitigate concurrency anomalies by issuing locking reads in
READ COMMITTED
transactions. These statements can block concurrent transactions that are issuing writes or other locking reads on the same rows.When using
READ COMMITTED
isolation, you do not need to implement client-side retries to handle serialization errors under transaction contention.READ COMMITTED
transactions never returnRETRY_SERIALIZABLE
errors, and will only return40001
errors in limited cases, as described in the following points.
READ COMMITTED
transactions can abort in certain scenarios:
Transactions at all isolation levels are subject to lock contention, where a transaction attempts to lock a row that is already locked by a write or locking read. In such cases, the later transaction is blocked until the earlier transaction commits or rolls back, thus releasing its lock on the row. Lock contention that produces a deadlock between two transactions will result in a transaction abort and a
40001
error (ABORT_REASON_ABORTED_RECORD_FOUND
orABORT_REASON_PUSHER_ABORTED
) returned to the client.Constraint violations will abort transactions at all isolation levels.
In rare cases under
READ COMMITTED
isolation, aRETRY_WRITE_TOO_OLD
orReadWithinUncertaintyIntervalError
error can be returned to the client if a statement has already begun streaming a partial result set back to the client and cannot retry transparently. By default, the result set is buffered up to 16 KiB before overflowing and being streamed to the client. You can configure the result buffer size using thesql.defaults.results_buffer.size
cluster setting.
Concurrency anomalies
Statements in concurrent READ COMMITTED
transactions can interleave with each other. This can create concurrency anomalies that are not permitted under SERIALIZABLE
isolation, which places concurrent transactions into a serializable ordering.
The behaviors described in this section assume the use of non-locking reads. You can prevent concurrency anomalies through the selective use of locking reads, which can also increase latency due to lock contention.
Non-repeatable reads and phantom reads
READ COMMITTED
transactions can serve different reads over the course of a transaction.
Non-repeatable reads return different row values because a concurrent transaction updated the values in between reads:
- Transaction
A
reads rowR
at timestamp1
. - Transaction
B
writes to rowR
and commits at timestamp2
. - Transaction
A
reads rowR
and gets a different result at timestamp3
.
Phantom reads return different rows because a concurrent transaction changed the set of rows that satisfy the row search:
- Transaction
A
reads the set of rowsS
at timestamp1
. - Transaction
B
inserts, deletes, or updates rows inS
and commits at timestamp2
. - Transaction
A
reads the set of rowsS
and gets a different result at timestamp3
.
Whereas statements in SERIALIZABLE
transactions see data that committed before the transaction began, statements in READ COMMITTED
transactions see data that committed before each statement began.
For details on how this is implemented, see Read snapshots.
Example: Non-repeatable reads and phantom reads
Session 1
Session 2
CREATE TABLE kv (k INT PRIMARY KEY, v INT);
INSERT INTO kv VALUES (1, 2);
Begin a READ COMMITTED
transaction and read a table row:
BEGIN TRANSACTION ISOLATION LEVEL READ COMMITTED;
SELECT * FROM kv WHERE v = 2;
k | v
----+----
1 | 2
READ COMMITTED
transaction:
BEGIN TRANSACTION ISOLATION LEVEL READ COMMITTED;
Update the table row, insert a new row, and commit the transaction:
UPDATE kv SET k = 2 WHERE v = 2;
INSERT INTO kv VALUES (3, 2);
COMMIT;
SELECT * FROM kv WHERE v = 2;
k | v
----+----
2 | 2
3 | 2
Lost update anomaly
The READ COMMITTED
conditions that permit non-repeatable reads and phantom reads also permit lost update anomalies, where an update from a transaction appears to be "lost" because it is overwritten by a concurrent transaction:
- Transaction
A
reads rowR
at timestamp1
. - Transaction
B
writes to rowR
and commits at timestamp2
. - Transaction
A
writes to rowR
and commits at timestamp3
.
The value of R
has changed while transaction A
is open. However, A
can still write to R
and commit, effectively overwriting the update from transaction B
.
Under SERIALIZABLE
isolation, transaction A
would have aborted with a RETRY_WRITE_TOO_OLD
error, prompting the client to retry the transaction.
Example: Lost update anomaly
Session 1
Session 2
CREATE TABLE kv (k INT PRIMARY KEY, v INT);
INSERT INTO kv VALUES (1, 2);
Begin a READ COMMITTED
transaction and read a table row:
BEGIN TRANSACTION ISOLATION LEVEL READ COMMITTED;
SELECT * FROM kv WHERE k = 1;
k | v
----+----
1 | 2
READ COMMITTED
transaction:
BEGIN TRANSACTION ISOLATION LEVEL READ COMMITTED;
Update the table row and commit the transaction:
UPDATE kv SET v = 3 WHERE k = 1;
COMMIT;
UPDATE kv SET v = 4 WHERE k = 1;
COMMIT;
Read the table row and see that it reflects the update from Session 1:
SELECT * FROM kv WHERE k = 1;
k | v
----+----
1 | 4
The update in Session 2 appears to be "lost" because its result is overwritten by a concurrent transaction. It is not lost at the database level, and can be found using AS OF SYSTEM TIME
and a timestamp earlier than the commit in Session 1:
SELECT * FROM kv AS OF SYSTEM TIME '2023-11-09 21:22:10' WHERE k = 1;
k | v
----+----
1 | 3
While concurrent READ COMMITTED
transactions can have their committed writes overwritten, uncommitted writes in READ COMMITTED
transactions cannot be overwritten.
Write skew anomaly
The following sequence of operations on a table is possible under READ COMMITTED
isolation:
- Transaction
A
reads rowR
at timestamp1
. - Transaction
B
reads rowS
at timestamp2
. - Transaction
A
writes to rowS
and commits at timestamp3
. - Transaction
B
writes to rowR
and commits at timestamp4
.
Transaction A
updates the value of S
based on the R
value it reads at timestamp 1
. Transaction B
updates the value of R
based on the S
value it reads at timestamp 2
. The value of S
has changed while transaction B
is open, but B
can still write and commit instead of aborting, since READ COMMITTED
transactions do not require serializability. This is the basis of potential write skew anomalies where two concurrent transactions each read values that the other subsequently updates.
For details on why this is allowed, see Read refreshing.
Example: Write skew anomaly
For an example of how a write skew anomaly can occur, see Demonstrate interleaved statements in READ COMMITTED
transactions.
Locking reads
To reduce the occurrence of concurrency anomalies in READ COMMITTED
isolation, you can strengthen the isolation of individual reads by using SELECT ... FOR UPDATE
or SELECT ... FOR SHARE
to issue locking reads on specific rows. Locking reads behave similarly to writes: they lock qualifying rows to prevent concurrent writes from modifying them until the transaction commits. Conversely, if a locking read finds that a row is exclusively locked by a concurrent transaction, it waits for the other transaction to commit or rollback before proceeding. A locking read in a transaction will always have the latest version of a row when the transaction commits.
The clause used with the SELECT
statement determines the lock strength of a locking read:
SELECT FOR UPDATE
obtains an exclusive lock on each qualifying row, blocking concurrent writes and locking reads on the row. Only one transaction can hold an exclusive lock on a row at a time, and only the transaction holding the exclusive lock can write to the row. For an example, see Reserve rows for updates using exclusive locks.SELECT FOR SHARE
obtains a shared lock on each qualifying row, blocking concurrent writes and exclusive locking reads on the row. Multiple transactions can hold a shared lock on a row at the same time. When multiple transactions hold a shared lock on a row, none can write to the row. A shared lock grants transactions mutual read-only access to a row, and ensures that they read the latest version of the row. For an example, see Reserve values using shared locks.
When a SELECT FOR UPDATE
or SELECT FOR SHARE
read is issued on a row, only the latest version of the row is returned to the client. Under READ COMMITTED
isolation, neither statement will block concurrent, non-locking reads.
When to use locking reads
Use locking reads in your application if certain READ COMMITTED
transactions must guarantee that the data they access will not be changed by intermediate writes.
Non-locking reads can allow intermediate writes to update rows before READ COMMITTED
transactions commit, potentially creating concurrency anomalies. Locking reads prevent such anomalies, but increase the amount of lock contention that may require intervention if latency becomes too high. Note that locking reads do not prevent phantom reads that are caused by the insertion of new rows, since only existing rows can be locked.
Locking reads are not effective for emulating SERIALIZABLE
transactions, which can avoid locking reads because they always retry or abort if reads are not current. As a result, READ COMMITTED
transactions that use locking reads will perform differently than SERIALIZABLE
transactions at various levels of concurrency.
To use locking reads:
If you need to read and later update a row within a transaction, use
SELECT ... FOR UPDATE
to acquire an exclusive lock on the row. This guarantees data integrity between the transaction's read and write operations.If you need to read the latest version of a row, and later update a different row within a transaction, use
SELECT ... FOR SHARE
to acquire a shared lock on the row. This blocks all concurrent writes on the row without unnecessarily blocking concurrent reads or otherSELECT ... FOR SHARE
queries.Tip:This allows an application to build cross-row consistency constraints by ensuring that rows that are read in a
READ COMMITTED
transaction will not change before the writes in the same transaction have been committed.
Examples
In this scenario:
- A hospital has an application for doctors to manage their on-call shifts.
- The hospital has a rule that at least one doctor must be on call at any one time.
- Two doctors are on call for a particular shift, and both of them try to request leave for the shift in two concurrent transactions.
- Under the
READ COMMITTED
isolation level, the write skew anomaly anomaly can potentially result in both doctors successfully booking leave and the hospital having no doctors on call for that particular shift.
The following examples demonstrate how to:
- Observe that
READ COMMITTED
transactions can serve different reads. - Use exclusive locks to strengthen isolation for
READ COMMITTED
transactions. - Use shared locks to reserve values in
READ COMMITTED
transactions.
Before you begin
Open the SQL shell using
cockroach demo
.Enable
READ COMMITTED
transactions:SET CLUSTER SETTING sql.txn.read_committed_isolation.enabled = 'true';
Create the
doctors
table:CREATE TABLE doctors ( id INT PRIMARY KEY, name TEXT );
Create the
schedules
table:CREATE TABLE schedules ( day DATE, doctor_id INT REFERENCES doctors (id), on_call BOOL, PRIMARY KEY (day, doctor_id) );
Add two doctors to the
doctors
table:INSERT INTO doctors VALUES (1, 'Abe'), (2, 'Betty');
Insert one week's worth of data into the
schedules
table:INSERT INTO schedules VALUES ('2023-12-01', 1, true), ('2023-12-01', 2, true), ('2023-12-02', 1, true), ('2023-12-02', 2, true), ('2023-12-03', 1, true), ('2023-12-03', 2, true), ('2023-12-04', 1, true), ('2023-12-04', 2, true), ('2023-12-05', 1, true), ('2023-12-05', 2, true), ('2023-12-06', 1, true), ('2023-12-06', 2, true), ('2023-12-07', 1, true), ('2023-12-07', 2, true);
Demonstrate interleaved statements in READ COMMITTED
transactions
Before proceeding, reset the example scenario:
UPDATE schedules SET on_call = true WHERE on_call = false;
Confirm that at least one doctor is on call each day of the week:
SELECT day, count(*) AS on_call FROM schedules
WHERE on_call = true
GROUP BY day
ORDER BY day;
day | on_call
-------------+----------
2023-12-01 | 2
2023-12-02 | 2
2023-12-03 | 2
2023-12-04 | 2
2023-12-05 | 2
2023-12-06 | 2
2023-12-07 | 2
Session 1
Session 2
2023-12-05
using the hospital's schedule management application.
Start a transaction:
BEGIN TRANSACTION ISOLATION LEVEL READ COMMITTED;
Check to make sure that another doctor is on call for 2023-12-05
:
SELECT * FROM schedules
WHERE day = '2023-12-05';
day | doctor_id | on_call
-------------+-----------+----------
2023-12-05 | 1 | t
2023-12-05 | 2 | t
In a new terminal (Session 2), open the SQL shell on your cockroach demo
cluster. Start a transaction:
BEGIN TRANSACTION ISOLATION LEVEL READ COMMITTED;
Check to make sure that another doctor is on call for 2023-12-05
:
SELECT * FROM schedules
WHERE day = '2023-12-05';
day | doctor_id | on_call
-------------+-----------+----------
2023-12-05 | 1 | t
2023-12-05 | 2 | t
2023-12-05
. Update the schedule to put Abe on leave:
UPDATE schedules SET on_call = false
WHERE day = '2023-12-05'
AND doctor_id = 1;
Read the rows for 2023-12-05
. Session 1 sees that only Abe is on leave once its transaction commits:
SELECT * FROM schedules
WHERE day = '2023-12-05';
day | doctor_id | on_call
-------------+-----------+----------
2023-12-05 | 1 | f
2023-12-05 | 2 | t
2023-12-05
. Update the schedule to put Betty on leave:
UPDATE schedules SET on_call = false
WHERE day = '2023-12-05'
AND doctor_id = 2;
Read the rows for 2023-12-05
. Session 2 sees that only Betty is on leave once its transaction commits:
SELECT * FROM schedules
WHERE day = '2023-12-05';
day | doctor_id | on_call
-------------+-----------+----------
2023-12-05 | 1 | t
2023-12-05 | 2 | f
COMMIT;
By design under READ COMMITTED
isolation, CockroachDB allows the transaction to commit even though its previous read (the SELECT
query) has changed due to the concurrent transaction in Session 2.
2023-12-05
again:
SELECT * FROM schedules
WHERE day = '2023-12-05';
day | doctor_id | on_call
-------------+-----------+----------
2023-12-05 | 1 | f
2023-12-05 | 2 | f
The result has changed because Session 1 committed earlier and updated the on_call
value for doctor 1, thus changing the read result for the transaction in Session 2.
If the transaction in Session 2 commits and updates the on_call
value for Betty, this will create a write skew anomaly. The result would be that neither Abe nor Betty is scheduled to be on call on 2023-12-05
.
Instead, the transaction should rollback so that the write skew anomaly does not commit:
ROLLBACK;
Reserve rows for updates using exclusive locks
Before proceeding, reset the example scenario:
UPDATE schedules SET on_call = true WHERE on_call = false;
Confirm that at least one doctor is on call each day of the week:
SELECT day, count(*) AS on_call FROM schedules
WHERE on_call = true
GROUP BY day
ORDER BY day;
day | on_call
-------------+----------
2023-12-01 | 2
2023-12-02 | 2
2023-12-03 | 2
2023-12-04 | 2
2023-12-05 | 2
2023-12-06 | 2
2023-12-07 | 2
Session 1
Session 2
2023-12-05
using the hospital's schedule management application.
Start a transaction:
BEGIN TRANSACTION ISOLATION LEVEL READ COMMITTED;
Check to make sure that another doctor is on call for 2023-12-05
. Use FOR UPDATE
to lock the rows so that only the current transaction can update them:
ORDER BY
clause to force locking to occur in a specific order. This prevents potential deadlock with another locking read on the same rows, which can cause the transaction to abort.
SELECT * FROM schedules
WHERE day = '2023-12-05'
ORDER BY doctor_id
FOR UPDATE;
day | doctor_id | on_call
-------------+-----------+----------
2023-12-05 | 1 | t
2023-12-05 | 2 | t
In a new terminal (Session 2), open the SQL shell on your cockroach demo
cluster. Start a transaction:
BEGIN TRANSACTION ISOLATION LEVEL READ COMMITTED;
Check to make sure that another doctor is on call for 2023-12-05
. Use FOR UPDATE
to lock the rows so that only the current transaction can update them:
SELECT * FROM schedules
WHERE day = '2023-12-05'
ORDER BY doctor_id
FOR UPDATE;
However, because Session 1 has already acquired an exclusive lock on these rows, the current transaction is blocked until Session 1 releases its lock.
2023-12-05
. Update the schedule to put Abe on leave:
UPDATE schedules SET on_call = false
WHERE day = '2023-12-05'
AND doctor_id = 1;
Commit the transaction:
COMMIT;
2023-12-05
, which show that Abe has already been put on leave for that day:
day | doctor_id | on_call
-------------+-----------+----------
2023-12-05 | 1 | f
2023-12-05 | 2 | t
Rollback the transaction:
ROLLBACK;
Reserve row values using shared locks
Before proceeding, reset the example scenario:
UPDATE schedules SET on_call = true WHERE on_call = false;
Confirm that at least one doctor is on call each day of the week:
SELECT day, count(*) AS on_call FROM schedules
WHERE on_call = true
GROUP BY day
ORDER BY day;
day | on_call
-------------+----------
2023-12-01 | 2
2023-12-02 | 2
2023-12-03 | 2
2023-12-04 | 2
2023-12-05 | 2
2023-12-06 | 2
2023-12-07 | 2
Session 1
Session 2
2023-12-05
using the hospital's schedule management application.
Start a transaction:
BEGIN TRANSACTION ISOLATION LEVEL READ COMMITTED;
Check to make sure that another doctor is on call for 2023-12-05
. Use FOR SHARE
to lock the rows so that they cannot be updated by another transaction:
SELECT * FROM schedules
WHERE day = '2023-12-05'
FOR SHARE;
day | doctor_id | on_call
-------------+-----------+----------
2023-12-05 | 1 | t
2023-12-05 | 2 | t
In a new terminal (Session 2), open the SQL shell on your cockroach demo
cluster. Start a transaction:
BEGIN TRANSACTION ISOLATION LEVEL READ COMMITTED;
Check to make sure that another doctor is on call for 2023-12-05
. Use FOR SHARE
to lock the rows so that they cannot be updated by another transaction:
SELECT * FROM schedules
WHERE day = '2023-12-05'
FOR SHARE;
day | doctor_id | on_call
-------------+-----------+----------
2023-12-05 | 1 | t
2023-12-05 | 2 | t
Shared locks are typically used when a transaction needs to read the latest version of a row, but does not need to update the row. With the rows locked by both Sessions 1 and 2, a third Session 3 is blocked from updating the rows:
UPDATE schedules SET on_call = false
WHERE day = '2023-12-05'
AND doctor_id = 1;
Once both Sessions 1 and 2 commit or rollback their transactions, Session 3 can complete the update to place Abe on leave:
UPDATE 1
COMMIT;
Read the rows for 2023-12-05
and confirm that Betty is still on call:
SELECT * FROM schedules
WHERE day = '2023-12-05';
day | doctor_id | on_call
-------------+-----------+----------
2023-12-05 | 1 | f
2023-12-05 | 2 | t
Known limitations
- Schema changes (e.g.,
CREATE TABLE
,CREATE SCHEMA
,CREATE INDEX
) cannot be performed within explicitREAD COMMITTED
transactions, and will cause transactions to abort. As a workaround, set the transaction's isolation level toSERIALIZABLE
. #114778 READ COMMITTED
transactions performingINSERT
,UPDATE
, orUPSERT
cannot accessREGIONAL BY ROW
tables in whichUNIQUE
andPRIMARY KEY
constraints exist, the region is not included in the constraint, and the region cannot be computed from the constraint columns.- Multi-column-family checks during updates are not supported under
READ COMMITTED
isolation. #112488 - Because locks acquired by foreign key checks,
SELECT FOR UPDATE
, andSELECT FOR SHARE
are fully replicated underREAD COMMITTED
isolation, some queries experience a delay for Raft replication. - Foreign key checks are not performed in parallel under
READ COMMITTED
isolation. SELECT FOR UPDATE
andSELECT FOR SHARE
statements are less optimized underREAD COMMITTED
isolation than underSERIALIZABLE
isolation. UnderREAD COMMITTED
isolation,SELECT FOR UPDATE
andSELECT FOR SHARE
usually perform an extra lookup join for every locked table when compared to the same queries underSERIALIZABLE
. In addition, some optimization steps (such as de-correlation of correlated subqueries) are not currently performed on these queries.- Regardless of isolation level,
SELECT FOR UPDATE
andSELECT FOR SHARE
statements in CockroachDB do not prevent insertion of new rows matching the search condition (i.e., phantom reads). This matches PostgreSQL behavior at all isolation levels. #120673