Data Replication

Replication is only supported for tables in the MergeTree family:

  • ReplicatedMergeTree
  • ReplicatedSummingMergeTree
  • ReplicatedReplacingMergeTree
  • ReplicatedAggregatingMergeTree
  • ReplicatedCollapsingMergeTree
  • ReplicatedVersionedCollapsingMergeTree
  • ReplicatedGraphiteMergeTree

Replication works at the level of an individual table, not the entire server. A server can store both replicated and non-replicated tables at the same time.

Replication does not depend on sharding. Each shard has its own independent replication.

Compressed data for INSERT and ALTER queries is replicated (for more information, see the documentation for ALTER).

CREATE, DROP, ATTACH, DETACH and RENAME queries are executed on a single server and are not replicated:

  • The CREATE TABLE query creates a new replicatable table on the server where the query is run. If this table already exists on other servers, it adds a new replica.
  • The DROP TABLE query deletes the replica located on the server where the query is run.
  • The RENAME query renames the table on one of the replicas. In other words, replicated tables can have different names on different replicas.

ClickHouse uses Apache ZooKeeper for storing replicas meta information. Use ZooKeeper version 3.4.5 or newer.

To use replication, set parameters in the zookeeper server configuration section.

Attention

Don’t neglect the security setting. ClickHouse supports the digest ACL scheme of the ZooKeeper security subsystem.

Example of setting the addresses of the ZooKeeper cluster:

  1. <zookeeper>
  2. <node index="1">
  3. <host>example1</host>
  4. <port>2181</port>
  5. </node>
  6. <node index="2">
  7. <host>example2</host>
  8. <port>2181</port>
  9. </node>
  10. <node index="3">
  11. <host>example3</host>
  12. <port>2181</port>
  13. </node>
  14. </zookeeper>

You can specify any existing ZooKeeper cluster and the system will use a directory on it for its own data (the directory is specified when creating a replicatable table).

If ZooKeeper isn’t set in the config file, you can’t create replicated tables, and any existing replicated tables will be read-only.

ZooKeeper is not used in SELECT queries because replication does not affect the performance of SELECT and queries run just as fast as they do for non-replicated tables. When querying distributed replicated tables, ClickHouse behavior is controlled by the settings max_replica_delay_for_distributed_queries and fallback_to_stale_replicas_for_distributed_queries.

For each INSERT query, approximately ten entries are added to ZooKeeper through several transactions. (To be more precise, this is for each inserted block of data; an INSERT query contains one block or one block per max_insert_block_size = 1048576 rows.) This leads to slightly longer latencies for INSERT compared to non-replicated tables. But if you follow the recommendations to insert data in batches of no more than one INSERT per second, it doesn’t create any problems. The entire ClickHouse cluster used for coordinating one ZooKeeper cluster has a total of several hundred INSERTs per second. The throughput on data inserts (the number of rows per second) is just as high as for non-replicated data.

For very large clusters, you can use different ZooKeeper clusters for different shards. However, this hasn’t proven necessary on the Yandex.Metrica cluster (approximately 300 servers).

Replication is asynchronous and multi-master. INSERT queries (as well as ALTER) can be sent to any available server. Data is inserted on the server where the query is run, and then it is copied to the other servers. Because it is asynchronous, recently inserted data appears on the other replicas with some latency. If part of the replicas are not available, the data is written when they become available. If a replica is available, the latency is the amount of time it takes to transfer the block of compressed data over the network.

By default, an INSERT query waits for confirmation of writing the data from only one replica. If the data was successfully written to only one replica and the server with this replica ceases to exist, the stored data will be lost. To enable getting confirmation of data writes from multiple replicas, use the insert_quorum option.

Each block of data is written atomically. The INSERT query is divided into blocks up to max_insert_block_size = 1048576 rows. In other words, if the INSERT query has less than 1048576 rows, it is made atomically.

Data blocks are deduplicated. For multiple writes of the same data block (data blocks of the same size containing the same rows in the same order), the block is only written once. The reason for this is in case of network failures when the client application doesn’t know if the data was written to the DB, so the INSERT query can simply be repeated. It doesn’t matter which replica INSERTs were sent to with identical data. INSERTs are idempotent. Deduplication parameters are controlled by merge_tree server settings.

During replication, only the source data to insert is transferred over the network. Further data transformation (merging) is coordinated and performed on all the replicas in the same way. This minimizes network usage, which means that replication works well when replicas reside in different datacenters. (Note that duplicating data in different datacenters is the main goal of replication.)

You can have any number of replicas of the same data. Yandex.Metrica uses double replication in production. Each server uses RAID-5 or RAID-6, and RAID-10 in some cases. This is a relatively reliable and convenient solution.

The system monitors data synchronicity on replicas and is able to recover after a failure. Failover is automatic (for small differences in data) or semi-automatic (when data differs too much, which may indicate a configuration error).

Creating Replicated Tables

The Replicated prefix is added to the table engine name. For example:ReplicatedMergeTree.

Replicated*MergeTree parameters

  • zoo_path — The path to the table in ZooKeeper.
  • replica_name — The replica name in ZooKeeper.

Example:

  1. CREATE TABLE table_name
  2. (
  3. EventDate DateTime,
  4. CounterID UInt32,
  5. UserID UInt32
  6. ) ENGINE = ReplicatedMergeTree('/clickhouse/tables/{layer}-{shard}/table_name', '{replica}')
  7. PARTITION BY toYYYYMM(EventDate)
  8. ORDER BY (CounterID, EventDate, intHash32(UserID))
  9. SAMPLE BY intHash32(UserID)

Example in deprecated syntax

  1. CREATE TABLE table_name
  2. (
  3. EventDate DateTime,
  4. CounterID UInt32,
  5. UserID UInt32
  6. ) ENGINE = ReplicatedMergeTree('/clickhouse/tables/{layer}-{shard}/table_name', '{replica}', EventDate, intHash32(UserID), (CounterID, EventDate, intHash32(UserID), EventTime), 8192)

As the example shows, these parameters can contain substitutions in curly brackets. The substituted values are taken from the ‘macros’ section of the configuration file. Example:

  1. <macros>
  2. <layer>05</layer>
  3. <shard>02</shard>
  4. <replica>example05-02-1.yandex.ru</replica>
  5. </macros>

The path to the table in ZooKeeper should be unique for each replicated table. Tables on different shards should have different paths.
In this case, the path consists of the following parts:

/clickhouse/tables/ is the common prefix. We recommend using exactly this one.

{layer}-{shard} is the shard identifier. In this example it consists of two parts, since the Yandex.Metrica cluster uses bi-level sharding. For most tasks, you can leave just the {shard} substitution, which will be expanded to the shard identifier.

table_name is the name of the node for the table in ZooKeeper. It is a good idea to make it the same as the table name. It is defined explicitly, because in contrast to the table name, it doesn’t change after a RENAME query.
HINT: you could add a database name in front of table_name as well. E.g. db_name.table_name

The replica name identifies different replicas of the same table. You can use the server name for this, as in the example. The name only needs to be unique within each shard.

You can define the parameters explicitly instead of using substitutions. This might be convenient for testing and for configuring small clusters. However, you can’t use distributed DDL queries (ON CLUSTER) in this case.

When working with large clusters, we recommend using substitutions because they reduce the probability of error.

Run the CREATE TABLE query on each replica. This query creates a new replicated table, or adds a new replica to an existing one.

If you add a new replica after the table already contains some data on other replicas, the data will be copied from the other replicas to the new one after running the query. In other words, the new replica syncs itself with the others.

To delete a replica, run DROP TABLE. However, only one replica is deleted – the one that resides on the server where you run the query.

Recovery After Failures

If ZooKeeper is unavailable when a server starts, replicated tables switch to read-only mode. The system periodically attempts to connect to ZooKeeper.

If ZooKeeper is unavailable during an INSERT, or an error occurs when interacting with ZooKeeper, an exception is thrown.

After connecting to ZooKeeper, the system checks whether the set of data in the local file system matches the expected set of data (ZooKeeper stores this information). If there are minor inconsistencies, the system resolves them by syncing data with the replicas.

If the system detects broken data parts (with the wrong size of files) or unrecognized parts (parts written to the file system but not recorded in ZooKeeper), it moves them to the detached subdirectory (they are not deleted). Any missing parts are copied from the replicas.

Note that ClickHouse does not perform any destructive actions such as automatically deleting a large amount of data.

When the server starts (or establishes a new session with ZooKeeper), it only checks the quantity and sizes of all files. If the file sizes match but bytes have been changed somewhere in the middle, this is not detected immediately, but only when attempting to read the data for a SELECT query. The query throws an exception about a non-matching checksum or size of a compressed block. In this case, data parts are added to the verification queue and copied from the replicas if necessary.

If the local set of data differs too much from the expected one, a safety mechanism is triggered. The server enters this in the log and refuses to launch. The reason for this is that this case may indicate a configuration error, such as if a replica on a shard was accidentally configured like a replica on a different shard. However, the thresholds for this mechanism are set fairly low, and this situation might occur during normal failure recovery. In this case, data is restored semi-automatically - by “pushing a button”.

To start recovery, create the node /path_to_table/replica_name/flags/force_restore_data in ZooKeeper with any content, or run the command to restore all replicated tables:

  1. sudo -u clickhouse touch /var/lib/clickhouse/flags/force_restore_data

Then restart the server. On start, the server deletes these flags and starts recovery.

Recovery After Complete Data Loss

If all data and metadata disappeared from one of the servers, follow these steps for recovery:

  1. Install ClickHouse on the server. Define substitutions correctly in the config file that contains the shard identifier and replicas, if you use them.
  2. If you had unreplicated tables that must be manually duplicated on the servers, copy their data from a replica (in the directory /var/lib/clickhouse/data/db_name/table_name/).
  3. Copy table definitions located in /var/lib/clickhouse/metadata/ from a replica. If a shard or replica identifier is defined explicitly in the table definitions, correct it so that it corresponds to this replica. (Alternatively, start the server and make all the ATTACH TABLE queries that should have been in the .sql files in /var/lib/clickhouse/metadata/.)
  4. To start recovery, create the ZooKeeper node /path_to_table/replica_name/flags/force_restore_data with any content, or run the command to restore all replicated tables: sudo -u clickhouse touch /var/lib/clickhouse/flags/force_restore_data

Then start the server (restart, if it is already running). Data will be downloaded from replicas.

An alternative recovery option is to delete information about the lost replica from ZooKeeper (/path_to_table/replica_name), then create the replica again as described in “Creating replicated tables“.

There is no restriction on network bandwidth during recovery. Keep this in mind if you are restoring many replicas at once.

Converting from MergeTree to ReplicatedMergeTree

We use the term MergeTree to refer to all table engines in the MergeTree family, the same as for ReplicatedMergeTree.

If you had a MergeTree table that was manually replicated, you can convert it to a replicated table. You might need to do this if you have already collected a large amount of data in a MergeTree table and now you want to enable replication.

If the data differs on various replicas, first sync it, or delete this data on all the replicas except one.

Rename the existing MergeTree table, then create a ReplicatedMergeTree table with the old name.
Move the data from the old table to the detached subdirectory inside the directory with the new table data (/var/lib/clickhouse/data/db_name/table_name/).
Then run ALTER TABLE ATTACH PARTITION on one of the replicas to add these data parts to the working set.

Converting from ReplicatedMergeTree to MergeTree

Create a MergeTree table with a different name. Move all the data from the directory with the ReplicatedMergeTree table data to the new table’s data directory. Then delete the ReplicatedMergeTree table and restart the server.

If you want to get rid of a ReplicatedMergeTree table without launching the server:

  • Delete the corresponding .sql file in the metadata directory (/var/lib/clickhouse/metadata/).
  • Delete the corresponding path in ZooKeeper (/path_to_table/replica_name).

After this, you can launch the server, create a MergeTree table, move the data to its directory, and then restart the server.

Recovery When Metadata in The ZooKeeper Cluster is Lost or Damaged

If the data in ZooKeeper was lost or damaged, you can save data by moving it to an unreplicated table as described above.