The SERIAL
pseudo data type is a keyword that can be used in lieu of a real data type when defining table columns. It is approximately equivalent to using an integer type with a DEFAULT
expression that generates different values every time it is evaluated. This default expression in turn ensures that inserts that do not specify this column will receive an automatically generated value instead of NULL
.
SERIAL
is provided only for compatibility with PostgreSQL. New applications should use real data types and a suitable DEFAULT
expression.
In most cases, we recommend using the UUID
data type with the gen_random_uuid()
function as the default value, which generates 128-bit values (larger than SERIAL
's maximum of 64 bits) and more uniformly scatters them across all of a table's underlying key-value ranges. UUIDs ensure more effectively that multiple nodes share the insert load when a UUID column is used in an index or primary key.
See this FAQ entry for more details.
Modes of operation
The keyword SERIAL
is recognized in CREATE TABLE
and is automatically translated to a real data type and a DEFAULT
expression during table creation. The result of this translation is then used internally by CockroachDB, and can be observed using SHOW CREATE
.
The chosen DEFAULT
expression ensures that different values are automatically generated for the column during row insertion. These are not guaranteed to increase monotonically, see this section below for details.
There are five possible translation modes for SERIAL
:
Mode | Description |
---|---|
rowid (default) |
SERIAL implies DEFAULT unique_rowid() . The real data type is always INT8 , regardless of the default_int_size session variable or sql.defaults.default_int_size cluster setting. |
unordered_rowid |
SERIAL implies DEFAULT unordered_unique_rowid() . The real data type is always INT8 , regardless of the default_int_size session variable or sql.defaults.default_int_size cluster setting. |
virtual_sequence |
SERIAL creates a virtual sequence and implies DEFAULT nextval(<seqname>) . The real data type is always INT8 , regardless of the default_int_size session variable or sql.defaults.default_int_size cluster setting. |
sql_sequence |
SERIAL creates a regular SQL sequence and implies DEFAULT nextval(<seqname>) . The real data type depends on SERIAL variant. |
sql_sequence_cached |
SERIAL creates a regular SQL sequence and implies DEFAULT nextval(<seqname>) , caching the results for reuse in the current session. The real data type depends on SERIAL variant. When the cache is empty, the underlying sequence will only be incremented once to populate the cache.When this mode is set, the sql.defaults.serial_sequences_cache_size cluster setting controls the number of values to cache in a user's session, with a default of 256. |
These modes can be configured with the session variable serial_normalization
.
The particular choice of DEFAULT
expression when clients use the SERIAL
keyword is subject to change in future versions of CockroachDB. Applications that wish to use unique_rowid()
specifically must use the full explicit syntax INT DEFAULT unique_rowid()
and avoid SERIAL
altogether.
Use ALTER ROLE ALL SET {sessionvar} = {val}
instead of the sql.defaults.*
cluster settings. This allows you to set a default value for all users for any session variable that applies during login, making the sql.defaults.*
cluster settings redundant.
Generated values for modes rowid
and virtual_sequence
In both modes rowid
and virtual_sequence
, a value is automatically generated using the unique_rowid()
function. The difference between rowid
and virtual_sequence
is that the latter setting also creates a virtual (pseudo) sequence in the database. However, in both cases the unique_rowid()
function is ultimately used to generate new values. This function produces a 64-bit integer (i.e., INT8
) from the current timestamp and ID of the node executing the INSERT
or UPSERT
operation. This behavior is statistically likely to be globally unique except in extreme cases (see this FAQ entry for more details).
The unique_rowid()
function creates a unique integer composed of the current time at a 10-microsecond granularity and the instance-id
. The instance-id
is stored in the lower 15 bits of the returned value and the timestamp is stored in the upper 48 bits. The top-most bit is left empty so that negative values are not returned. The 48-bit timestamp field provides for 89 years of timestamps. CockroachDB uses a custom epoch (Jan 1, 2015) in order to utilize the entire timestamp range.
Also, because value generation using unique_rowid()
does not require inter-node coordination, it is much faster than the other mode sql_sequence
discussed below when multiple SQL clients are writing to the table from different nodes.
Values generated by the unique_rowid()
function do not respect the default_int_size
session variable or sql.defaults.default_int_size
cluster setting.
Generated values for mode sql_sequence
and sql_sequence_cached
In both modes, a regular SQL sequence is automatically created alongside the table where SERIAL
is specified.
The actual data type is determined as follows:
SERIAL variant |
Real data type |
---|---|
SERIAL2 , SMALLSERIAL |
INT2 |
SERIAL4 |
INT4 |
SERIAL |
INT |
SERIAL8 , BIGSERIAL |
INT8 |
Every insert or upsert into the table will then use nextval()
to increment the sequence and produce increasing values.
Because SQL sequences persist the current sequence value in the database, inter-node coordination is required when multiple clients use the sequence concurrently via different nodes. This can cause contention and impact performance negatively.
Therefore, applications should consider using unique_rowid()
or gen_random_uuid()
as discussed in this FAQ entry instead of sequences when possible.
Note that sql_sequence_cached
will perform fewer distributed calls to increment sequences, resulting in better performance than sql_sequence
. However, cached sequences may result in large gaps between serial sequence numbers if a session terminates before using all the values in its cache.
Generated values for mode unordered_rowid
In this mode, a value is generated using the unordered_unique_rowid()
function. This function is similar to the rowid
mode but does not guarantee ordering.
Auto-incrementing is not always sequential
It's a common misconception that the auto-incrementing types in PostgreSQL and MySQL generate strictly sequential values. However, there can be gaps and the order is not completely guaranteed:
- Each insert increases the sequence by one, even when the insert is not committed. This means that auto-incrementing types may leave gaps in a sequence.
- Two concurrent transactions can commit in a different order than their use of sequences, and thus "observe" the values to decrease relative to each other. This effect is amplified by automatic transaction retries.
These are fundamental properties of a transactional system with non-transactional sequences. PostgreSQL, MySQL, and CockroachDB do not increase sequences transactionally with other SQL statements, so these effects can happen in any case.
To experience this for yourself, run through the following example in PostgreSQL:
Create a table with a
SERIAL
column:> CREATE TABLE increment (a SERIAL PRIMARY KEY);
Run four transactions for inserting rows:
> BEGIN; > INSERT INTO increment DEFAULT VALUES; > ROLLBACK;
> BEGIN; > INSERT INTO increment DEFAULT VALUES; > COMMIT;
> BEGIN; > INSERT INTO increment DEFAULT VALUES; > ROLLBACK;
> BEGIN; > INSERT INTO increment DEFAULT VALUES; > COMMIT;
View the rows created:
> SELECT * from increment;
a ---------------------- 645994218934534145 645994218978082817 (2 rows)
Since each insert increased the sequence in column
a
by one, the first committed insert got the value645994218934534145
, and the second committed insert got the value645994218978082817
. As you can see, the values aren't strictly sequential, and the last value doesn't give an accurate count of rows in the table.
In summary, the SERIAL
type in PostgreSQL and CockroachDB, and the AUTO_INCREMENT
type in MySQL, all behave the same in that they do not create strict sequences. CockroachDB will likely create more gaps than these other databases, but will generate these values much faster. An alternative is to use SEQUENCE
.
Additional scenarios
If two transactions occur concurrently, CockroachDB cannot guarantee monotonically increasing IDs (i.e., first commit is smaller than second commit). Here are three more scenarios that demonstrate this:
Scenario 1
- At time 1, transaction
T1
BEGIN
s. - At time 2, transaction
T2
BEGIN
s on the same node (from a different client). - At time 3, transaction
T1
creates aSERIAL
value,x
. - At time 3 + 2 microseconds, transaction
T2
creates aSERIAL
value,y
. - At time 4, transaction
T1
COMMIT
s. - At time 5, transaction
T2
COMMIT
s.
If this happens, CockroachDB cannot guarantee whether x < y
or x > y
, despite the fact T1
and T2
began and were committed in different times. In this particular example, it's even likely that x = y
because there is less than a 10-microsecond difference and the SERIAL
values are constructed from the number of microseconds in the current time.
Scenario 2
- At time 1, transaction
T1
BEGIN
s. - At time 1, transaction
T2
BEGIN
s somewhere else, on a different node. - At time 2, transaction
T1
creates aSERIAL
value,x
. - At time 3, transaction
T2
creates aSERIAL
value,y
. - At time 4, transaction
T1
COMMIT
s. - At time 4, transaction
T2
COMMIT
s.
If this happens, CockroachDB cannot guarantee whether x < y
or x > y
. Both can happen, even though the transactions began and committed at the same time. However it's sure that x != y
because the values were generated on different nodes.
Scenario 3
- At time 1, transaction
T1
BEGIN
s. - At time 2, transaction
T1
creates aSERIAL
value,x
. - At time 3, transaction
T1
COMMIT
s. - At time 4, transaction
T2
BEGIN
s somewhere else, on a different node. - At time 5, transaction
T2
creates aSERIAL
value,y
. - At time 6, transaction
T2
COMMIT
s.
There is less than a 250-microsecond difference between the system clocks of the two nodes.
If this happens, CockroachDB cannot guarantee whether x < y
or x > y
. Even though the transactions "clearly" occurred one "after" the other, perhaps there was a clock skew between the two nodes and the system time of the second node is set earlier than the first node.
Example
Use SERIAL
to auto-generate primary keys
In this example, we create a table with the SERIAL
column as the primary key so we can auto-generate unique IDs on insert.
> CREATE TABLE serial (a SERIAL PRIMARY KEY, b STRING, c BOOL);
The SHOW COLUMNS
statement shows that the SERIAL
type is just an alias for INT
with unique_rowid()
as the default.
> SHOW COLUMNS FROM serial;
column_name | data_type | is_nullable | column_default | generation_expression | indices | is_hidden
--------------+-----------+-------------+----------------+-----------------------+-----------+------------
a | INT8 | false | unique_rowid() | | {primary} | false
b | STRING | true | NULL | | {} | false
c | BOOL | true | NULL | | {} | false
(3 rows)
When we insert rows without values in column a
and display the new rows, we see that each row has defaulted to a unique value in column a
.
> INSERT INTO serial (b,c) VALUES ('red', true), ('yellow', false), ('pink', true);
> INSERT INTO serial (a,b,c) VALUES (123, 'white', false);
> SELECT * FROM serial;
a | b | c
---------------------+--------+--------
123 | white | false
645993277462937601 | red | true
645993277462970369 | yellow | false
645993277463003137 | pink | true
(4 rows)