RelStorage Options

Specify these options in zope.conf, as parameters for the constructor, or as attributes of a relstorage.options.Options instance. In the latter two cases, use underscores instead of dashes in the option names.

General Settings


The name of the storage.

Defaults to a descriptive name that includes the adapter connection parameters, except the database password.


If true, only reads may be executed against the storage.

The default is false.


If this option is set to true (the default), the adapter will create and use a history-preserving database schema (like FileStorage). A history-preserving schema supports ZODB-level undo, but also grows more quickly and requires extensive packing on a regular basis.


Using ZODB’s time-travel feature to get historical connections ( will result in RDBMS connections being left open in a transaction until they are removed from the ZODB historical connection pool and closed. PostgreSQL refers to this state as “idle in transaction”, and it ties up certain MVCC resources on the RDBMS server.

If this option is set to false, the adapter will create and use a history-free database schema. Undo will not be supported, but the database will not grow as quickly. The database will still require regular garbage collection (which is accessible through the database pack mechanism.)


This option must not change once the database schema has been installed, because the schemas for history-preserving and history-free storage are different. RelStorage will refuse to initialize if it detects this value has been altered.

If you want to convert between a history-preserving and a history-free database, use the Copying Data Between ZODB Storages: zodbconvert utility to copy to a new database.


During commit, RelStorage acquires a database-wide lock. This option specifies how long to wait for the lock before failing the attempt to commit. The default is 30 seconds.

Changed in version 2.0.0b1: Add support for lock timeouts on PostgreSQL 9.3 and above. Previously no PostgreSQL version supported timeouts.

Changed in version 3.0a5: With the implementation of ZODB 5’s parallel commit feature, this option has changed slightly. Now, individual rows will be locked instead of locking entire tables or using advisory locks. This timeout now controls the locking of individual rows. Because rows are locked in specific orders, in extreme cases of overlapping transactions, a transaction might wait almost this length of time multiple times. However, it will always be making forward progress and will still raise an exception if it cannot make forward progress after this amount of time.

There is still a single row locked by all transactions, but that is done after most commit work has been accomplished and is hopefully fast. The exception is if shared-blob-dir is true (deprecated), in which case the single row is locked much earlier.


During commit, RelStorage acquires a database-wide lock. This option specifies the lock ID.

This option currently applies only to the Oracle adapter.


Normally, RelStorage will create or update the database schema on start-up.

Set this option to false if you need to connect to a RelStorage database without automatic creation or updates.



If supplied, the storage will provide ZODB blob support; this option specifies the name of the directory to hold blob data. The directory will be created if it does not exist.

If no value (or an empty value) is provided, then no blob support will be provided.


When this option is false (the default, and recommended), the blob directory is treated as a cache. It should be on a local filesystem that properly supports file locks. It may be shared among clients on that same machine. Blob data is stored safely in the relational database, and the blob directory holds a cache of blobs.

If true (not recommended), the blob directory is assumed to be shared among all clients using NFS or similar (file locks must also be supported); blob data will be stored only on the filesystem and not in the database. It is critical to have backups of this directory, as this is the only source of blobs.


When this option is true, the ability to do parallel commits is reduced. It is highly recommended to set this value to false. The possible exception is SQLite, which doesn’t support parallel commit anyway.

Changed in version 3.0a7: The default changed from true to false.


Maximum size of the blob cache, in bytes. If not set (the default), the cache size isn’t checked and the blob directory will grow without bounds. This should be either empty or significantly larger than the largest blob you store. At least 1 gigabyte is recommended for typical databases. More is recommended if you store large files such as videos, CD/DVD images, or virtual machine images.

When configured, the size is checked when a process opens a storage for the first time, and at intervals based on blob-cache-size-check.

This option allows suffixes such as “mb” or “gb”.

This option is ignored if shared-blob-dir is true.


Blob cache check size as percent of blob-cache-size: “10” means “10%”. The blob cache size will be checked when this many bytes have been loaded into the cache by any one process. Defaults to 10% of the blob cache size.

This option is ignored if shared-blob-dir is true.


When this value is false (the default), the blob cache size will be checked using an internal native thread (even when the process has been monkey-patched with gevent: gevent’s threadpool will be used). When this value is true, an external subprocess will be used to check the size.

For large blob caches, where checking the size takes measurable time, using an external process may improve response time for the application by reducing contention for the GIL. It may also be helpful for gevent applications.

This is not recommended on Windows, where opening a file from multiple processes can be a problem.


When ZODB blobs are stored in MySQL, RelStorage breaks them into chunks to minimize the impact on RAM. This option specifies the chunk size for new blobs. If RAM is available and the network connection to the database server is fast, a larger value can be more efficient because it will result in fewer roundtrips to the server.


On MySQL, this value cannot exceed the server’s max_allowed_packet setting. If blob chunks are larger than that, it won’t be possible to upload them. If blob chunks are uploaded and then that setting is later reduced, it won’t be possible to download blobs that exceed that value.

The driver may also influence this.

On PostgreSQL and Oracle, this value is used as the memory buffer size for blob upload and download operations.

The default is 1048576 (1 megabyte). This option allows suffixes such as “mb” or “gb”.

This option has no effect if shared-blob-dir is true (because blobs are not stored on the server).



If this option is provided, it specifies a text file that contains a list of database replicas the adapter can choose from. For MySQL and PostgreSQL, put in the replica file a list of host:port or host values, one per line. For Oracle, put in a list of DSN values. Blank lines and lines starting with # are ignored.

The adapter prefers the first replica specified in the file. If the first is not available, the adapter automatically tries the rest of the replicas, in order. If the file changes, the adapter will drop existing SQL database connections and make new connections when ZODB starts a new transaction.


Like the replica-conf option, but the referenced text file provides a list of database replicas to use only for read-only load connections. This allows RelStorage to load objects from read-only database replicas, while using read-write replicas for all other database interactions.

If this option is not provided, load connections will fall back to the replica pool specified by replica-conf. If ro-replica-conf is provided but replica-conf is not, RelStorage will use replicas for load connections but not for other database interactions.

Note that if read-only replicas are asynchronous, the next interaction after a write operation might not be up to date. When that happens, RelStorage will log a “backward time travel” warning and clear the ZODB cache to prevent consistency errors. This will likely result in temporarily reduced performance as the ZODB cache is repopulated.

New in version 1.6.0.


If this option has a nonzero value, when the adapter selects a replica other than the primary replica, the adapter will try to revert to the primary replica after the specified timeout (in seconds).

The default is 600, meaning 10 minutes.


Specifies what to do when a database connection is stale. This is especially applicable to asynchronously replicated databases: RelStorage could switch to a replica that is not yet up to date.

When revert-when-stale is false (the default) and the database connection is stale, RelStorage will raise a ReadConflictError if the application tries to read or write anything. The application should react to the ReadConflictError by retrying the transaction after a delay (possibly multiple times.) Once the database catches up, a subsequent transaction will see the update and the ReadConflictError will not occur again.

When revert-when-stale is true and the database connection is stale, RelStorage will log a warning, clear the affected ZODB connection cache (to prevent consistency errors), and let the application continue with database state from an earlier transaction. This behavior is intended to be useful for highly available, read-only ZODB clients. Enabling this option on ZODB clients that read and write the database is likely to cause confusion for users whose changes seem to be temporarily reverted.

New in version 1.6.0.

GC and Packing


If pack-gc is false, pack operations do not perform garbage collection. Garbage collection is enabled by default.

If garbage collection is disabled, pack operations keep at least one revision of every object that hasn’t been deleted. With garbage collection disabled, the pack code does not need to follow object references, making packing conceivably much faster. However, some of that benefit may be lost due to an ever increasing number of unused objects.

Disabling garbage collection is required in a multi-database to prevent breaking inter-database references. The only safe way to collect and then pack databases in a multi-database is to use zc.zodbdgc and run multi-zodb-gc, and only then pack each individual database.


In history-free databases, packing after running multi-zodb-gc is not necessary. The garbage collection process itself handles the packing. Packing is only required in history-preserving databases.

Changed in version 3.0: Add support for zc.zodbdgc to history-preserving databases.

Objects that have been deleted will be removed during a pack with pack-gc disabled.

Changed in version 2.0: Add support for zc.zodbdgc to history-free databases.


If pack-prepack-only is true, pack operations perform a full analysis of what to pack, but no data is actually removed. After a pre-pack, the pack_object, pack_state, and pack_state_tid tables are filled with the list of object states and objects that would have been removed.

If pack-gc is true, the object_ref table will also be fully populated. The object_ref table can be queried to discover references between stored objects.


If pack-skip-prepack is true, the pre-pack phase is skipped and it is assumed the pack_object, pack_state and pack_state_tid tables have been filled already. Thus packing will only affect records already targeted for packing by a previous pre-pack analysis run.

Use this option together with pack-prepack-only to split packing into distinct phases, where each phase can be run during different timeslots, or where a pre-pack analysis is run on a copy of the database to alleviate a production database load.


Packing occurs in batches of transactions; this specifies the timeout in seconds for each batch. Note that some database configurations have unpredictable I/O performance and might stall much longer than the timeout.

The default timeout is 1.0 seconds.


Changed in version 3.0a5: This option is now deprecated and does nothing. The commit lock is not held during packing anymore. The remainder of the documentation for this option only applies to older versions.

Before each pack batch, the commit lock is requested. If the lock is already held by for a regular commit, packing is paused for a short while. This option specifies how long the pack process should be paused before attempting to get the commit lock again.

The default delay is 5.0 seconds.

Database Caching

In addition to the ZODB Connection object caches, RelStorage uses pickle caches to reduce the number of queries to the database server. (This is similar to ZEO.) Caches can be both local to a process (within its memory) and remote (and shared between many processes) but outside of special circumstances this is not recommended.

These options affect all caching operations.


The prefix for all keys in the remote shared cache; also used as part of persistent cache file names. All clients using a database should use the same cache-prefix. Defaults to the database name. (For example, in PostgreSQL, the database name is determined by executing SELECT current_database().) Set this if you have multiple databases with the same name.

Changed in version 1.6.0b1: Start defaulting to the database name.

Local Caching

RelStorage caches pickled objects in memory, similar to a ZEO cache. The “local” cache is shared between all threads in a process. An adequately sized local cache is important for the highest possible performance. Using a suitably-sized local cache, especially with persistent data files, can make a substantial performance difference, even if the write volume is relatively high.

For help understanding and tuning the cache behaviour, see Client Cache Tracing.


This option configures the approximate maximum amount of memory the cache should consume, in megabytes. It defaults to 10.

Set to 0 to disable the in-memory cache. (This is not recommended.)


This option configures the maximum size of an object’s pickle (in bytes) that can qualify for the “local” cache. The size is measured after compression. Larger objects can still qualify for the remote cache.

The default is 16384 (1 << 14) bytes.

Changed in version 2.0b2: Measure the size after compression instead of before.


This option configures compression within the “local” cache. This option names a Python module that provides two functions, compress() and decompress(). Supported values include zlib, bz2, and none (no compression).

The default is none to avoid copying data more than necessary.

If you use the compressing storage wrapper zc.zlibstorage, this option automatically does nothing. With other compressing storage wrappers this should be set to none.

New in version 1.6.


This is an advanced option related to the MVCC implementation used by RelStorage’s cache.

The default is 100,000 on CPython, 50,000 on PyPy.

Changed in version 3.0a9: This parameter has changed meanings. Previously, it controlled a memory allocation per-ZODB connection, and it also controlled how often each connection would have to make an expensive database query. Smaller values of around 20,000 are appropriate on these older versions.

In 3.0a9 this is a global value shared among all connections. It is no longer directly related to polling, but if a read connection is open for more new or changed objects than this limit, the next time the connection is used it will drop its ZODB object cache rather than attempt to pick through the invalidations. Idle connections like that may indicate a ZODB connection pool that’s too large; consider enabling the idle connection timeout.

Changed in version 3.0a3: Increase the sizes again. With better persistent caching, these become increasingly important.

Changed in version 2.0b7: Double the default size from 10000 to 20000 on CPython. The use of LLBTree for the internal data structure means we use much less memory than we did before.

Persistent Local Caching

Like ZEO, RelStorage can store its local cache to disk for a quicker application warmup period.

Changed in version 3.0a7: Raise a descriptive error if the sqlite3 library is too old to be used by the local cache. RelStorage requires at least 3.8.3 but works better with 3.15 and best with 3.24 or newer.

This feature is no longer considered experimental.

Changed in version 3.0a1: The persistent file format and contents have been substantially changed to produce much better hit rates.

Reading and writing the cache files is slower, however, as the cache size gets larger.

The cache files must be located on a local (not network) filesystem.

New in version 2.0b2: This is a new, experimental feature. While there should be no problems enabling it, the exact details of its function are subject to change in the future based on feedback and experience.


The path to a directory where the local cache will be saved when the database is closed. On startup, RelStorage will look in this directory for cache files to load into memory.

This option can dramatically reduce the amount of time it takes for your application to warm up after a restart, especially if there were relatively few writes in the meantime. Some synthetic benchmarks show an 8-10x improvement after a restart.

This is similar to the ZEO persistent cache, but adds no overhead to normal transactions.

This directory will be populated with files written each time a RelStorage instance is closed. If multiple RelStorage processes are working against the same database (for example, the workers of gunicorn), then they will each write and read files in this directory. On startup, the files will be combined to get the “warmest” possible cache.

The time taken to load the cache file (which only occurs when RelStorage is first opened) and the time taken to save the cache file (which only occurs when the database is closed) are proportional to the total size of the cache; thus, a cache that is too large (holding many unused entries) will slow down startup/shutdown for no benefit. However, the timing is probably not a practical concern compared to the disk usage; on one system, a 300MB uncompressed cache file can be saved in 3-4 seconds and read in 2-3 seconds.

This directory can be shared among storages connected to different databases, so long as they all have a distinct cache-prefix.

If this directory does not exist, we will attempt to create it on startup. This directory must be a local filesystem, not network storage.


If the database (ZODB.DB object) is not properly closed, then the cache files will not be written.


This requires at least sqlite3 3.8.3 or better. Improved performance comes with version 3.15 and the best performance is with 3.24 or higher.

Deprecated Options

The following local cache options apply only to RelStorage 2.x. They are deprecated, ignored, and have no impact on RelStorage 3.0a1 or newer. RelStorage releases in 2021 will no longer accept them.


How many files that cache-local-dir will be allowed to contain before files start getting reused. Set this equal to the number of workers that will be sharing the directory.

The default is 20.


Whether to compress the persistent cache files on disk. The default is false because individual entries are usually already compressed and the tested workloads do not show a space benefit from the compression; in addition, compression can slow the reading and writing process by 2 to 3 times or more (and hence slow down opening the storage).

New in version 2.0b5.


The maximum number of files to read to populate the cache on startup.

By default, RelStorage will read all the appropriate files (so up to cache-local-dir-count files), from newest to oldest, collecting the distinct entries out of them, until the cache is fully populated (cache-local-mb) or there are no more files. This ensures that after startup, all workers have the most fully populated cache. This strategy works well if the workers have a good distribution of work (relatively few overlapping items) and the cache size is relatively small; after startup they will all be equally warm without spending too much startup time.

However, if the workers all do nearly the same work (so most items in the cache files are the same) and the cache sizes are very large, then the benefits of reading each subsequent file diminish (because it has very few if any new entries to add, and reading them all takes a lot of time). In that case, set this value to 1 to only read the first (“best”) cache file.

For situations in between, choose a number in between.

New in version 2.0b5.


The approximate maximum size of each individual cache file on disk. When not specified (the default), the maximum file size will be the same as cache-local-mb.

This is an approximate number because there is some overhead associated with the storage format that varies based on the number of entries in the cache.

RelStorage will write to disk, from most important to least important, the entries in the cache until all the entries are written or this limit is reached. If you use a size smaller than cache-local-mb, however, you may miss important entries that are only used at application startup.

New in version 2.0b7.

Remote Caching

RelStorage can use Memcached servers as a secondary, semi-persistent database cache. This is generally not recommended; most RelStorage queries are simple enough that even in extremely large databases Memcached is not faster than the database. The need for a cache that persists across restarts is better met with local persistent caches.

However, if local persistent caches are not an option, memcache may be useful if the ratio of writes to reads is extremely low (because they add substantial overhead to each write operation), and if the database server is behind a high-latency connection or otherwise responds slowly.


Remote caching is scheduled for removal with RelStorage releases in 2021.


Specifies a list of memcached servers. Using memcached with RelStorage improves the speed of frequent object accesses while slightly reducing the speed of other operations.

Provide a list of host:port pairs, separated by whitespace. “” is a common setting. Some memcached modules, such as pylibmc, allow you to specify a path to a Unix socket instead of a host:port pair.

The default is to disable memcached integration.

New in version 1.1rc1.


Specifies which Python memcache module to use. The default is “relstorage.pylibmc_wrapper”, which requires pylibmc. An alternative module is “memcache”, a pure Python module. If you use the memcache module, use at least version 1.47.

This option has no effect unless cache-servers is set.