A beginner’s guide to read and write skew phenomena

Introduction

In my article about ACID and database transactions, I introduced the three phenomena described by the SQL standard:

  • dirty read
  • non-repeatable read
  • phantom read

While these are good to differentiate the four isolation levels (Read Uncommitted, Read Committed, Repeatable Read and Serializable), in reality, there are more phenomena to take into consideration as well. The 1995 paper (A Critique of ANSI SQL Isolation Levels) introduces the other phenomena that are omitted from the standard specification.

In my High-Performance Java Persistence book, I decided to insist on the Transaction chapter as it is very important for both data access effectiveness and efficiency.

Domain model

For the following examples, I’m going to use the following two entities:

ReadWriteSkew

In our fictional application, when the Post title is changed, the author must be recorded in the associated PostDetails record.
If read and write skew are not prevented, this domain model constraint can be compromised, as you will see in the following test cases.

Read skew

The following test simulates how a read skew can happen:

doInConnection(aliceConnection -> {
    prepareConnection(aliceConnection);
    String title = selectStringColumn(
        aliceConnection, 
        selectPostTitleSql()
    );
    executeSync(() -> {
        doInConnection(bobConnection -> {
            prepareConnection(bobConnection);
            try {
                update(
                    bobConnection, 
                    updatePostTitleParamSql(), 
                    new Object[]{"Bob"}
                );
                update(
                    bobConnection, 
                    updatePostDetailsAuthorParamSql(), 
                    new Object[]{"Bob"}
                );
            } catch (Exception e) {
                LOGGER.info("Exception thrown", e);
                preventedByLocking.set(true);
            }
        });
    });
    String createdBy = selectStringColumn(
        aliceConnection, 
        selectPostDetailsAuthorSql()
    );
});
  • Alice selects a Post title
  • Bob sneaks in and updates the Post and the PostDetails entities
  • Alice thread is resumed and she selects the PostDetails record

If read skew is permitted, Alice sees Bob’s update and she can assume that the previous Post version (that she read at the beginning of her transaction) was issued by Bob (which might not be accurate).

Running this test on the four most common relation database systems gives the following results:

Database isolation level Read skew
Oracle Read Committed Yes
Oracle Serializable No
SQL Server Read Uncommitted Yes
SQL Server Read Committed Yes
SQL Server Read Committed Snapshot Isolation Yes
SQL Server Repeatable Read No
SQL Server Serializable No
SQL Server Snapshot Isolation No
PostgreSQL Read Uncommitted Yes
PostgreSQL Read Committed Yes
PostgreSQL Repeatable Read No
PostgreSQL Serializable No
MySQL Read Uncommitted Yes
MySQL Read Committed Yes
MySQL Repeatable Read No
MySQL Serializable No

Write skew

To emulate write skew, you need to execute the following test case:

doInConnection(aliceConnection -> {
    prepareConnection(aliceConnection);
    String title = selectStringColumn(
        aliceConnection, 
        selectPostTitleSql()
    );
    String createdBy = selectStringColumn(
        aliceConnection, 
        selectPostDetailsAuthorSql()
    );
    executeSync(() -> {
        doInConnection(bobConnection -> {
            prepareConnection(bobConnection);
            try {
                String bobTitle = selectStringColumn(
                    bobConnection, 
                    selectPostTitleSql()
                );
                String bonCreatedBy = selectStringColumn(
                    bobConnection, 
                    selectPostDetailsAuthorSql()
                );
                update(
                    bobConnection, 
                    updatePostTitleParamSql(), 
                    new Object[]{"Bob"}
                );
            } catch (Exception e) {
                LOGGER.info("Exception thrown", e);
                preventedByLocking.set(true);
            }
        });
    });
    update(
        aliceConnection, 
        updatePostDetailsAuthorParamSql(), 
        new Object[]{"Alice"}
    );
});
  • Alice selects the Post title and the author from the PostDetails record
  • Bob also selects the Post title and the associated author, but he decides to update the title only
  • Alice thinks of updating the PostDetails record without changing the Post title

If write skew is permitted, Alice and Bob disjoint writes will be executed without being blocked by the constraint that governs both records.

Running this test on the four most common relation database systems gives the following results:

Database isolation level Write skew
Oracle Read Committed Yes
Oracle Serializable Yes
SQL Server Read Uncommitted Yes
SQL Server Read Committed Yes
SQL Server Read Committed Snapshot Isolation Yes
SQL Server Repeatable Read No
SQL Server Serializable No
SQL Server Snapshot Isolation Yes
PostgreSQL Read Uncommitted Yes
PostgreSQL Read Committed Yes
PostgreSQL Repeatable Read Yes
PostgreSQL Serializable No
MySQL Read Uncommitted Yes
MySQL Read Committed Yes
MySQL Repeatable Read Yes
MySQL Serializable No

If you enjoyed this article, I bet you are going to love my book as well.

Conclusion

  • Write skew is prevalent among Multi-Version Concurrency Control mechanisms and Oracle cannot prevent it even when claiming to be using Serializable, which in fact is just a Snapshot Isolation level instead.
  • SQL Server default locking-based isolation levels can prevent write skews when using Repeatable Read and Serializable. Neither one of the snapshot isolation levels (MVCC-based) can prevent/detect it instead.
  • PostgreSQL prevents it using its more advanced Serializable Snapshot Isolation level
  • MySQL employs shared locks when using Serializable, so the write skew can be prevented even if InnoDB is also MVCC-based

Enter your email address to follow this blog and receive notifications of new posts by email.

Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s