Database Management Systems uses MVCC to avoid the problem of Writers blocking Readers and vice-versa, by making use of multiple versions of data. There are essentially two approaches to multi-version concurrency.
Approaches for MVCC
The first approach is to store multiple versions of records in the database, and garbage collect records when they are no longer required. This is the approach adopted by PostgreSQL and Firebird/Interbase. SQL Server also uses asomewhat similar approach with the difference that old versions are stored in tempdb (a database different from the main database). The second approach is to keep only the latest version of data in the database, but reconstruct older versions of data dynamically as required by using undo. This is approach was adopted by Oracle and MySQL/InnoDB.
MVCC in PostgreSQL
In PostgreSQL, when a row is updated, a new version (called a tuple) of the row is created and inserted into the table. The previous version is provided as a pointer to the new version. The previous version is marked “expired", but remains in the database until it is “garbage collected.” In order to support multi-versioning, each tuple has additional data recorded with it:
- xmin - The ID of the transaction that inserted/updated the row and created this tuple.
- xmax - The transaction that deleted the row, or created a new version of this tuple. Initially this field is null.
Transaction status is maintained in CLOG which resides in $Data/pg_clog. This table contains two bits of status information for each transaction; the possible states are in-progress, committed, or
aborted. PostgreSQL does not undo changes to database rows when a transaction aborts - it simply marks the transaction as aborted in CLOG. A PostgreSQL table therefore may contain data from aborted transactions.
Enterprise-ready Postgres tools for high availability, monitoring, and disaster recovery. Download Now.
A Vacuum cleaner process is provided to garbage collect expired/aborted versions of a row. The Vacuum Cleaner also deletes index entries associated with tuples that are garbage collected. A tuple is visible if it’s xmin is valid and xmax is not. “Valid" means “either committed or the current transaction". To avoid consulting the CLOG table repeatedly, PostgreSQL maintains status flags in the tuple that indicate whether the tuple is “known committed" or “known aborted".
MVCC in Oracle
Oracle maintains old versions in rollback segments (also known as 'undo log'). A transaction ID is not a sequential number; instead, it is made of a set of numbers that points to the transaction entry (slot) in a Rollback segment header. Rollback segments have the property that new transactions can reuse storage and transaction slots used by older transactions that are committed or aborted. This automatic reuse facility enables Oracle to manage large numbers of transactions using a finite set of rollback segments.
The header block of the rollback segment is used as a transaction table. Here the status of a transaction is maintained (called System Change Number, or SCN, in Oracle). Rather than storing a transaction ID with each row in the page, Oracle saves space by maintaining an array of unique transaction IDs separately within the page, and stores only the offset of this array with the row. Along with each transaction ID, Oracle stores a pointer to the last undo record created by the transaction for the page. Not only are table rows stored in this way, Oracle employs the same techniques when storing index rows. This is one of the major differences between PostgreSQL and Oracle.
When an Oracle transaction starts, it makes a note of the current SCN. When reading a table or an index page, Oracle uses the SCN number to determine if the page contains the effects of transactions that should not be visible to the current transaction. Oracle checks the commit status of a transaction by looking up the associated Rollback segment header, but, to save time, the first time a transaction is looked up, its status is recorded in the page itself to avoid future lookups. If the page is found to contain the effects of invisible transactions, then Oracle recreates an older version of the page by undoing the effects of each such transaction. It scans the undo records associated with each transaction and applies them to the page until the effects of those transactions are removed. The new page created this way is then used to access the tuples within it.
Record Header in Oracle
A row header never grows, always a fixed size. For non-cluster tables, the row header is 3 bytes. One byte is used to store flags, one byte to indicate if the row is locked (for example because it's updated but not committed), and one byte for the column count.
MVCC in SQL Server
Snapshot isolation and read committed using row versioning are enabled at the database level. Only databases that require this option must enable it and incur the overhead associated with it. Versioning effectively starts with a copy-on-write mechanism that is invoked when a row is modified or deleted. Row versioning–based transactions can effectively "view" the consistent version of the data from these previous row versions.
Row versions are stored within the version store that is housed within the tempdb database. More specifically, when a record in a table or index is modified, the new record is stamped with the "sequence_number" of the transaction that is performing the modification. The old version of the record is copied to the version store, and the new record contains a pointer to the old record in the version store. If multiple long-running transactions exist and multiple "versions" are required, records in the version store might contain pointers to even earlier versions of the row.
Version store cleanup in SQL Server
SQL Server manages the version store size automatically, and maintains a cleanup thread to make sure it does not keep versioned rows around longer than needed. For queries running under Snapshot Isolation, the version store retains the row versions until the transaction that modified the data completes and the transactions containing any statements that reference the modified data complete. For SELECT statements running under Read Committed Snapshot Isolation, a particular row version is no longer required, and is removed, once the SELECT statement has executed.
If tempdb actually runs out of free space, SQL Server calls the cleanup function and will increase the size of the files, assuming we configured the files for auto-grow. If the disk gets so full that the files cannot grow, SQL Server will stop generating versions. If that happens, any snapshot query that needs to read a version that was not generated due to space constraints will fail.
Record Header in SQL Server
4 bytes long
- two bytes of record metadata (record type)
- two bytes pointing forward in the record to the NULL bitmap. This is offset to some actual data in record (fixed length columns).
Versioning tag - this is a 14-byte structure that contains a timestamp plus a pointer into the version store in tempdb. Here timestamp is trasaction_seq_number, the only time that rows get versioning info added to record is when it’s needed to support a versioning operation.
As the versioning information is optional, I think that is the reason they could store this info in index records as well without much impact.
Conclusion of Study
As other databases store version/visibility information in index, that makes index cleanup easier (as it is no longer tied to heap for visibility information). The advantage for not storing the visibility information in index is that for Delete operations, we don't need to perform an index delete and probably the size of index record could be somewhat smaller. Oracle and probably MySQL (Innodb) needs to write the record in undo segment for Insert statement whereas in PostgreSQL/SQL Server, the new record version is created only when a row is modified or deleted.
Only changed values are written to undo whereas PostgreSQL/SQL Server creates a complete new tuple for modified row. This avoids bloat in the main heap segment. Both Oracle and SQL Server has some way to restrict the growth of version information whereas PostgreSQL/PPAS doesn't have any way.
Amit Kapila emphasizes database internals and is a technical team leader at EDB. A 19-year database veteran, Amit has developed deep expertise in PostgreSQL, Oracle®, and in-memory databases. He participates actively in developing PostgreSQL and reviewing new features and is also a Committer and a Major Developer in PostgreSQL. Amit's major work in PostgreSQL includes parallel-query, performance improvements for multi-core machines, scalability and durable hash-indexes. In the past, he has worked on integrating the in-memory storage engine to PostgreSQL-based code and also improved the Oracle performance by doing statement caching.