This page describes limitations we identified in the CockroachDB v1.0 release. For limitations that have been subsequently resolved, this page also calls out the release incuding the change.
Removing all rows from large tables
When removing all rows from a table via a TRUNCATE
statement or a DELETE
statement without a WHERE
clause, CockroachDB batches the entire operation as a single transaction. For large tables, this can cause the nodes containing the table data to either crash or exhibit poor performance due to elevated memory and CPU usage.
As a workaround, when you need to remove all rows from a large table:
- Use
SHOW CREATE TABLE
to get the table schema. - Use
DROP TABLE
to remove the table. - Use
CREATE TABLE
with the output from step 1 to recreate the table.
Schema changes within transactions
Within a single transaction:
- DDL statements cannot be mixed with DML statements. As a workaround, you can split the statements into separate transactions.
- A
CREATE TABLE
statement containingFOREIGN KEY
orINTERLEAVE
clauses cannot be followed by statements that reference the new table. This also applies to runningTRUNCATE
on such a table becauseTRUNCATE
implicitly drops and recreates the table. - A table cannot be dropped and then recreated with the same name. This is not possible within a single transaction because
DROP TABLE
does not immediately drop the name of the table. As a workaround, split theDROP TABLE
andCREATE TABLE
statements into separate transactions.
Schema changes between executions of prepared statements
When the schema of a table targeted by a prepared statement changes after the prepared statement is created, future executions of the prepared statement could result in an error. For example, adding a column to a table referenced in a prepared statement with a SELECT *
clause will result in an error:
CREATE TABLE users (id INT PRIMARY KEY);
PREPARE prep1 AS SELECT * FROM users;
ALTER TABLE users ADD COLUMN name STRING;
INSERT INTO users VALUES (1, 'Max Roach');
EXECUTE prep1;
ERROR: cached plan must not change result type
SQLSTATE: 0A000
It's therefore recommended to explicitly list result columns instead of using SELECT *
in prepared statements, when possible.
Join flags when restoring a backup onto new machines
In our deployment tutorials, when starting the first node of a cluster, the --join
flag should be empty, but when starting all subsequent nodes, the --join
flag should be set to the address of node 1. This approach ensures that all nodes have access to a copy of the first key-value range, which is part of a meta-index identifying where all range replicas are stored, and which nodes require to initialize themselves and start accepting incoming connections.
Ensuring that all nodes have access to a copy of the first key-value range is more difficult when restoring from a whole-cluster backup onto machines with different IP addresses than the original cluster. In this case, the --join
flags must form a fully-connected directed graph. The easiest way to do this is to put all of the new nodes' addresses into each node's --join
flag, which ensures all nodes can join a node with a copy of the first key-value range.
INSERT ON CONFLICT
vs. UPSERT
When inserting/updating all columns of a table, and the table has no secondary indexes, we recommend using an UPSERT
statement instead of the equivalent INSERT ON CONFLICT
statement. Whereas INSERT ON CONFLICT
always performs a read to determine the necessary writes, the UPSERT
statement writes without reading, making it faster.
This issue is particularly relevant when using a simple SQL table of two columns to simulate direct KV access. In this case, be sure to use the UPSERT
statement.
Repeated or combined commands in the SQL shell history
Our built-in SQL shell stores previously executed commands in the shell's history. In some cases, these commands are unexpectedly duplicated.
Also, in some terminals, such as st
or xterm
without tmux
, previously executed commands are combined into a single command in the SQL shell history.
Using \|
to perform a large input in the SQL shell
In the built-in SQL shell, using the \|
operator to perform a large number of inputs from a file can cause the server to close the connection. This is because \|
sends the entire file as a single query to the server, which can exceed the upper bound on the size of a packet the server can accept from any client (16MB).
As a workaround, execute the file from the command line with cat data.sql | cockroach sql
instead of from within the interactive shell.
New values generated by DEFAULT
expressions during ALTER TABLE ADD COLUMN
When executing an ALTER TABLE ADD COLUMN
statement with a DEFAULT
expression, new values generated:
- use the default search path regardless of the search path configured in the current session via
SET SEARCH_PATH
. - use the UTC time zone regardless of the time zone configured in the current session via
SET TIME ZONE
. - have no default database regardless of the default database configured in the current session via
SET DATABASE
, so you must specify the database of any tables they reference. - use the transaction timestamp for the
statement_timestamp()
function regardless of the time at which theALTER
statement was issued.
Load-based lease rebalancing in uneven latency deployments
When nodes are started with the --locality
flag, CockroachDB attempts to place the replica lease holder (the replica that client requests are forwarded to) on the node closest to the source of the request. This means as client requests move geographically, so too does the replica lease holder.
However, you might see increased latency caused by a consistently high rate of lease transfers between datacenters in the following case:
- Your cluster runs in datacenters which are very different distances away from each other.
- Each node was started with a single tier of
--locality
, e.g.,--locality=datacenter=a
. - Most client requests get sent to a single datacenter because that's where all your application traffic is.
To detect if this is happening, open the Admin UI, select the Queues dashboard, hover over the Replication Queue graph, and check the Leases Transferred / second data point. If the value is consistently larger than 0, you should consider stopping and restarting each node with additional tiers of locality to improve request latency.
For example, let's say that latency is 10ms from nodes in datacenter A to nodes in datacenter B but is 100ms from nodes in datacenter A to nodes in datacenter C. To ensure A's and B's relative proximity is factored into lease holder rebalancing, you could restart the nodes in datacenter A and B with a common region, --locality=region=foo,datacenter=a
and --locality=region=foo,datacenter=b
, while restarting nodes in datacenter C with a different region, --locality=region=bar,datacenter=c
.
Roundtrip to STRING
does not respect precedences of :::
and -
Queries with constant expressions that evaluate to 2**-63 might get incorrectly rejected, for example:
> CREATE TABLE t (i int PRIMARY KEY);
> INSERT INTO t VALUES (1), (2), (3);
> SELECT (-9223372036854775808) ::: INT;
> SELECT i FROM t WHERE (i, i) < (1, -9223372036854775808);
pq: ($0, $0) < (1, - 9223372036854775808:::INT): tuples ($0, $0), (1, - 9223372036854775808:::INT) are not comparable at index 2: numeric constant out of int64 range
Overload resolution for collated strings
Many string operations are not properly overloaded for collated strings, for example:
> SELECT 'string1' || 'string2';
+------------------------+
| 'string1' || 'string2' |
+------------------------+
| string1string2 |
+------------------------+
(1 row)
> SELECT ('string1' collate en) || ('string2' collate en);
pq: unsupported binary operator: <collatedstring{en}> || <collatedstring{en}>
Quoting collation locales containing uppercase letters
Quoting a collation locale containing uppercase letters results in an error, for example:
> CREATE TABLE a (b STRING COLLATE "DE");
invalid syntax: statement ignored: invalid locale "DE": language: tag is not well-formed at or near ")"
CREATE TABLE a (b STRING COLLATE "DE");
^
As a workaround, make the locale lowercase or remove the quotes, for example:
> CREATE TABLE a (b STRING COLLATE "de");
> CREATE TABLE b (c STRING COLLATE DE);
Creating views with array types
Because arrays are not supported, attempting to create a view with an array in the SELECT
query crashes the node that receives the request.
Dropping a database containing views
When a view queries multiple tables or a single table multiple times (e.g., via UNION
), dropping the
database containing the tables fails silently.
Qualifying a column that comes from a view
It is not possible to fully qualify a column that comes from a view because the view gets replaced by an anonymous subquery, for example:
> CREATE TABLE test (a INT, b INT);
> CREATE VIEW Caps AS SELECT a, b FROM test;
> SELECT sum(Caps.a) FROM Caps GROUP BY b;
pq: source name "caps" not found in FROM clause
Write and update limits for a single transaction
A single transaction can contain at most 100,000 write operations (e.g., changes to individual columns) and at most 64MiB of combined updates. When a transaction exceeds these limits, it gets aborted. INSERT INTO .... SELECT FROM ...
queries commonly encounter these limits.
If you need to increase these limits, you can update the cluster-wide settings kv.transaction.max_intents
and kv.raft.command.max_size
. For INSERT INTO .. SELECT FROM
queries in particular, another workaround is to manually page through the data you want to insert using separate transactions.
Max size of a single column family
When creating or updating a row, if the combined size of all values in a single column family exceeds the max range size (64MiB by default) for the table, the operation may fail, or cluster performance may suffer.
As a workaround, you can either manually split a table's columns into multiple column families, or you can create a table-specific zone configuration with an increased max range size.
Simultaneous client connections and running queries on a single node
When a node has both a high number of client connections and running queries, the node may crash due to memory exhaustion. This is due to CockroachDB not accurately limiting the number of clients and queries based on the amount of available RAM on the node.
To prevent memory exhaustion, monitor each node's memory usage and ensure there is some margin between maximum CockroachDB memory usage and available system RAM. For more details about memory usage in CockroachDB, see this blog post.
SQL subexpressions and memory usage
Many SQL subexpressions (e.g., ORDER BY
, UNION
/INTERSECT
/EXCEPT
, GROUP BY
, subqueries) accumulate intermediate results in RAM on the node processing the query. If the operator attempts to process more rows than can fit into RAM, the node will either crash or report a memory capacity error. For more details about memory usage in CockroachDB, see this blog post.
Counting distinct rows in a table
When using count(DISTINCT a.*)
to count distinct rows in a table based on a subset of the columns, as opposed to count(*)
, the results are almost always incorrect, for example:
> CREATE TABLE t (a INT, b INT);
> INSERT INTO t VALUES (1, 2), (1, 3), (2, 1);
> SELECT count(DISTINCT t.*) FROM t;
+---------------------+
| count(DISTINCT t.*) |
+---------------------+
| 1 |
+---------------------+
(1 row)
As a workaround, list the columns explicitly, for example:
> SELECT count(DISTINCT (t.a, t.b)) FROM t;
+----------------------------+
| count(DISTINCT (t.a, t.b)) |
+----------------------------+
| 3 |
+----------------------------+
(1 row)
Running on Windows as a non-admin user
By default, CockroachDB periodically rotates the file it writes logs to, as well as a symlink pointing to the file it's currently using. However, on Windows, non-admin users cannot create symlinks, which prevents CockroachDB from starting because it cannot create logs.
To resolve this issue, non-admin users must log to stdout
(instead of files) by passing --log-dir=
(with the empty value) to the cockroach start
command, for example:
$ cockroach.exe start --log-dir= --insecure
Query planning for OR
expressions
Given a query like SELECT * FROM foo WHERE a > 1 OR b > 2
, even if there are appropriate indexes to satisfy both a > 1
and b > 2
, the query planner performs a full table or index scan because it cannot use both conditions at once.
Privileges for DELETE
and UPDATE
Every DELETE
or UPDATE
statement constructs a SELECT
statement, even when no WHERE
clause is involved. As a result, the user executing DELETE
or UPDATE
requires both the DELETE
and SELECT
or UPDATE
and SELECT
privileges on the table.
Dropping an index interleaved into another index on the same table
In the unlikely case that you interleave an index into another index on the same table and then drop the interleaved index, future DDL operations on the table will fail.
For example:
> CREATE TABLE t1 (id1 INT PRIMARY KEY, id2 INT, id3 INT);
> CREATE INDEX c ON t1 (id2)
STORING (id1, id3)
INTERLEAVE IN PARENT t1 (id2);
> SHOW INDEXES FROM t1;
+-------+---------+--------+-----+--------+-----------+---------+----------+
| Table | Name | Unique | Seq | Column | Direction | Storing | Implicit |
+-------+---------+--------+-----+--------+-----------+---------+----------+
| t1 | primary | true | 1 | id1 | ASC | false | false |
| t1 | c | false | 1 | id2 | ASC | false | false |
| t1 | c | false | 2 | id1 | N/A | true | false |
| t1 | c | false | 3 | id3 | N/A | true | false |
+-------+---------+--------+-----+--------+-----------+---------+----------+
(4 rows)
> DROP INDEX t1@c;
> DROP TABLE t1;
pq: invalid interleave backreference table=t1 index=3: index-id "3" does not exist
> TRUNCATE TABLE t1;
pq: invalid interleave backreference table=t1 index=3: index-id "3" does not exist
> ALTER TABLE t1 RENAME COLUMN id3 TO id4;
pq: invalid interleave backreference table=t1 index=3: index-id "3" does not exist
Order of dumped schemas and incorrect schemas of dumped views
When using the cockroach dump
command to export the schemas of all tables and views in a database, the schemas are ordered alphabetically by name. This is not always an ordering in which the tables and views can be successfully recreated. Also, the schemas of views are dumped incorrectly as CREATE TABLE
statements.
For example, consider a database test
with 2 tables and 1 view. Table a
has a foreign key reference to table c
, and view b
references table c
:
> CREATE DATABASE test;
> CREATE TABLE test.c (a INT PRIMARY KEY, b STRING);
> CREATE TABLE test.a (a INT PRIMARY KEY, b INT NOT NULL REFERENCES test.c (a))
> CREATE VIEW test.b AS SELECT b FROM test.c;
When you dump the schemas of the tables and views in database test
, they are ordered alphabetically by name, and the schema for view b
is incorrectly listed as a CREATE TABLE
statement:
$ cockroach dump --insecure --dump-mode=schema > dump.txt
$ cat dump.txt
CREATE TABLE a (
a INT NOT NULL,
b INT NOT NULL,
CONSTRAINT "primary" PRIMARY KEY (a ASC),
CONSTRAINT fk_b_ref_c FOREIGN KEY (b) REFERENCES c (a),
FAMILY "primary" (a, b)
);
CREATE TABLE b (
b STRING NOT NULL
);
CREATE TABLE c (
a INT NOT NULL,
b STRING NULL,
CONSTRAINT "primary" PRIMARY KEY (a ASC),
FAMILY "primary" (a, b)
);
If you tried to import these incorrectly ordered schemas to restore the test
database or create a new database, the import would fail.
As a workaround, before using exported schemas to recreate tables and views, you must reorder the schemas so that tables with foreign keys and views are listed after the tables they reference, and you must fix CREATE
statements for views:
CREATE TABLE c (
a INT NOT NULL,
b STRING NULL,
CONSTRAINT "primary" PRIMARY KEY (a ASC),
FAMILY "primary" (a, b)
);
CREATE TABLE a (
a INT NOT NULL,
b INT NOT NULL,
CONSTRAINT "primary" PRIMARY KEY (a ASC),
CONSTRAINT fk_b_ref_c FOREIGN KEY (b) REFERENCES c (a),
FAMILY "primary" (a, b)
);
CREATE VIEW b AS SELECT b FROM c;
Dumping data for a view
When using the cockroach dump
command to export the data of a view, the dump fails. This is because, unlike standard tables, a view does not contain any physical data; instead, it is a stored SELECT
query that, when requested, dynamically forms a virtual table.
For example, consider a database test
with a standard table t1
and a view v1
that references t1
:
> CREATE DATABASE test;
> CREATE TABLE test.t1 (a INT PRIMARY KEY, b STRING);
> INSERT INTO test.t1 VALUES (1, 'a'), (2, 'b');
> CREATE VIEW test.v1 AS SELECT b FROM test.t1;
Trying to dump the data of the view results in an error:
$ cockroach dump test v1 --insecure --dump-mode=data
Error: pq: column name "rowid" not found
Failed running "dump"
This error occurs when trying to dump all the data in database test
as well:
$ cockroach dump test --insecure --dump-mode=data
INSERT INTO t1 (a, b) VALUES
(1, 'a'),
(2, 'b');
Error: pq: column name "rowid" not found
Failed running "dump"
As a workound, when dumping all the data in a database, explicitly list the tables that should be dumped, excluding any views:
$ cockroach dump test t1 --insecure --dump-mode=data
INSERT INTO t1 (a, b) VALUES
(1, 'a'),
(2, 'b');