Delta Lake provides ACID transaction guarantees between reads and writes. This means that:
- Multiple writers across multiple clusters can simultaneously modify a table partition and see a consistent snapshot view of the table and there will be a serial order for these writes.
- Readers will continue to see the consistent snapshot view of the table that the <DBROrSpark> job started with, even when the table is modified during the job.
Delta Lake uses optimistic concurrency control to provide transactional guarantees between writes. Under this mechanism, writes operate in three stages:
- Read: Reads (if needed) the latest available version of the table to identify which files need to be modified (that is, rewritten).
- Write: Stages all the changes by writing new data files.
- Validate and commit: Before committing the changes, checks whether the proposed changes conflict with any other changes that may have been concurrently committed since the snapshot that was read. If there are no conflicts, all the staged changes are committed as a new versioned snapshot, and the write operation succeeds. However, if there are conflicts, the write operation fails with a concurrent modification exception rather than corrupting the table as would happen with the write operation on a Parquet table.
The isolation level of a table defines the degree to which a transaction must be isolated from modifications made by concurrent operations. For information on the isolation levels supported by Delta Lake on Databricks, see Isolation levels.
The following table describes which pairs of write operations can conflict in each isolation level.
|INSERT||UPDATE, DELETE, MERGE INTO||OPTIMIZE|
|UPDATE, DELETE, MERGE INTO||Can conflict in Serializable, cannot conflict in WriteSerializable||Can conflict in Serializable and WriteSerializable|
|OPTIMIZE||Cannot conflict||Can conflict in Serializable and WriteSerializable||Can conflict in Serializable and WriteSerializable|
In all cases marked “can conflict”, whether the two operations will conflict depends on whether they operate on the same set of files. You can make the two sets of files disjoint by partitioning the table by the same columns as those used in the conditions of the operations. For example, the two commands
UPDATE table WHERE date > '2010-01-01' ... and
DELETE table WHERE date < '2010-01-01' will conflict if the table is not partitioned by date, as both can attempt to modify the same set of files. Partitioning the table by
date will avoid the conflict. Hence, partitioning a table according to the conditions commonly used on the command can reduce conflicts significantly. However, partitioning a table by a column that has high cardinality can lead to other performance issues due to large number of subdirectories.
When a transaction conflict occurs, you will observe one of the following exceptions.
This exception occurs when a concurrent operation adds files in the same partition (or anywhere in an unpartitioned table) that your operation reads. The file additions can be caused by
With the default isolation level of
WriteSerializable, files added by blind
INSERT operations (that is, operations that blindly append data without reading any data) do not conflict with any operation, even if they touch the same partition (or anywhere in an unpartitioned table). If the isolation level is set to
Serializable, then blind appends may conflict.
This exception is often thrown during concurrent
MERGE operations. While the concurrent operations may be physically updating dufferent partition directories, one of them may read the same partition that the other one concurrently updates, thus causing a conflict. You can avoid this by making the separation explicit in the operation condition. Consider the following example.
// Target 'deltaTable' is partitioned by date and country deltaTable.as("t").merge( source.as("s"), "s.user_id = t.user_id AND s.date = t.date AND s.country = t.country") .whenMatched().updateAll() .whenNotMatched().insertAll() .execute()
Suppose you run the above code concurrently for different dates or countries. Since each job is working on an independent partition on the target Delta table, you don’t expect any conflicts. However, the condition is not explicit enough and can scan the entire table and can conflict with concurrent operations updating any other partitions. Instead, you can rewrite your statement to add specific date and country to the merge condition, as shown below.
// Target 'deltaTable' is partitioned by date and country deltaTable.as("t").merge( source.as("s"), "s.user_id = t.user_id AND s.date = t.date AND s.country = t.country AND t.date = '" + <date> + "' AND t.country = '" + <country> + "'") .whenMatched().updateAll() .whenNotMatched().insertAll() .execute()
This operation is now safe to run concurrently on different dates and countries.
This exception occurs when a concurrent operation deleted a file that your operation read. Common causes are a
MERGE operation that rewrites files.
This exception occurs when a concurrent operation deleted a file that your operation also deletes. This could be caused by two concurrent <compaction> operations rewriting the same files.
This exception occurs when a concurrent transaction updates the metadata of a Delta table. Common causes are
ALTER TABLE operations or writes to your Delta table that update the schema of the table.
If a streaming query using the same checkpoint location is started multiple times concurrently and tries to write to the Delta table at the same time. You should never have two streaming queries use the same checkpoint location and run at the same time.
This occurs when your Delta table is upgraded to a new version. For future operations to succeed you may need to upgrade your Delta Lake version.
See Table versioning for more details.