Cache

This filter was introduced in MariaDB MaxScale 2.1.

Overview

Note that the cache is still experimental and that non-backward compatible changes may be made.

The cache filter is a simple cache that is capable of caching the result of SELECTs, so that subsequent identical SELECTs are served directly by MaxScale, without the queries being routed to any server.

The cache will be used and populated in the following circumstances: There is no explicit transaction active, that is, autocommit is used, there is an explicitly read-only transaction (that is,START TRANSACTION READ ONLY) active, or * there is a transaction active and no statement that modifies the database has been performed.

In practice, the last bullet point basically means that if a transaction has been started with BEGIN, START TRANSACTION or START TRANSACTION READ WRITE, then the cache will be used and populated until the first UPDATE, INSERT or DELETE statement is encountered.

By default, it is ensured that the cache is not used in the following circumstances:

  • The SELECT uses any of the following functions: BENCHMARK, CONNECTION_ID, CONVERT_TZ, CURDATE, CURRENT_DATE, CURRENT_TIMESTAMP, CURTIME, DATABASE, ENCRYPT, FOUND_ROWS, GET_LOCK, IS_FREE_LOCK, IS_USED_LOCK, LAST_INSERT_ID, LOAD_FILE, LOCALTIME, LOCALTIMESTAMP, MASTER_POS_WAIT, NOW, RAND, RELEASE_LOCK, SESSION_USER, SLEEP, SYSDATE, SYSTEM_USER, UNIX_TIMESTAMP, USER, UUID, UUID_SHORT.
  • The SELECT accesses any of the following fields: CURRENT_DATE, CURRENT_TIMESTAMP, LOCALTIME, LOCALTIMESTAMP
  • The SELECT uses system or user variables.

In order to ensure that, all SELECT statements have to be parsed, which carries a significant performance cost. If it is known that there are no such statements or that it does not matter even if they are cached, that safety measure can be turned off. Please read performance for more details.

Limitations

All of these limitations may be addressed in forthcoming releases.

Invalidation

Currently there is no cache invalidation, apart from time-to-live.

Prepared Statements

Resultsets of prepared statements are not cached.

Security

The cache is not aware of grants.

The implication is that unless the cache has been explicitly configured who the caching should apply to, the presence of the cache may provide a user with access to data he should not have access to.

Please read the section Security for more detailed information.

Configuration

The cache is simple to add to any existing service. However, some experimentation may be required in order to find the configuration settings that provide the maximum benefit.

[Cache]
type=filter
module=cache
hard_ttl=30
soft_ttl=20
rules=...
...

[Cached Routing Service]
type=service
...
filters=Cache

Each configured cache filter uses a storage of its own. That is, if there are two services, each configured with a specific cache filter, then, even if queries target the very same servers the cached data will not be shared.

Two services can use the same cache filter, but then either the services should use the very same servers or a completely different set of servers, where the used table names are different. Otherwise there can be unintended sharing.

Filter Parameters

The cache filter has no mandatory parameters but a range of optional ones. Note that it is advisable to specify max_size to prevent the cache from using up all memory there is, in case there is very litte overlap among the queries.

storage

The name of the module that provides the storage for the cache. That module will be loaded and provided with the value of storage_options as argument. For instance:

storage=storage_inmemory

The default is storage_inmemory.

See Storage for what storage modules are available.

storage_options

A comma separated list of arguments to be provided to the storage module, specified in storage, when it is loaded. Note that the needed arguments depend upon the specific module. For instance,

storage_options=storage_specific_option1=value1,storage_specific_option2=value2

hard_ttl

Hard time to live; the maximum amount of time - in seconds - the cached result is used before it is discarded and the result is fetched from the backend (and cached). See also soft_ttl below.

hard_ttl=60

The default value is 0, which means no limit.

soft_ttl

Soft time to live; the amount of time - in seconds - the cached result is used before it is refreshed from the server. When soft_ttl has passed, the result will be refreshed when the first client requests the value.

However, as long as hard_ttl has not passed, all other clients requesting the same value will use the result from the cache while it is being fetched from the backend. That is, as long as soft_ttl but not hard_ttl has passed, even if several clients request the same value at the same time, there will be just one request to the backend.

soft_ttl=60

The default value is 0, which means no limit. If the value of soft_ttl is larger than hard_ttl it will be adjusted down to the same value.

max_resultset_rows

Specifies the maximum number of rows a resultset can have in order to be stored in the cache. A resultset larger than this, will not be stored.

max_resultset_rows=1000

The default value is 0, which means no limit.

max_resultset_size

Specifies the maximum size of a resultset, for it to be stored in the cache. A resultset larger than this, will not be stored. The size can be specified as described here.

max_resultset_size=128Ki

The default value is 0, which means no limit.

Note that the value of max_resultset_size should not be larger than the value of max_size.

max_count

The maximum number of items the cache may contain. If the limit has been reached and a new item should be stored, then an older item will be evicted.

Note that if cached_data is thread_specific then this limit will be applied to each cache separately. That is, if a thread specific cache is used, then the total number of cached items is #threads * the value of max_count.

max_count=1000

The default value is 0, which means no limit.

max_size

The maximum size the cache may occupy. If the limit has been reached and a new item should be stored, then some older item(s) will be evicted to make space. The size can be specified as described here.

Note that if cached_data is thread_specific then this limit will be applied to each cache separately. That is, if a thread specific cache is used, then the total size is #threads * the value of max_size.

max_size=100Mi

The default value is 0, which means no limit.

rules

Specifies the path of the file where the caching rules are stored. A relative path is interpreted relative to the data directory of MariaDB MaxScale.

rules=/path/to/rules-file

cached_data

An enumeration option specifying how data is shared between threads. The allowed values are:

  • shared: The cached data is shared between threads. On the one hand it implies that there will be synchronization between threads, on the other hand that all threads will use data fetched by any thread.
  • thread_specific: The cached data is specific to a thread. On the one hand it implies that no synchonization is needed between threads, on the other hand that the very same data may be fetched and stored multiple times.
cached_data=thread_specific

Default is shared. See max_count and max_size what implication changing this setting to thread_specific has.

selects

An enumeration option specifying what approach the cache should take with respect to SELECT statements. The allowed values are:

  • assume_cacheable: The cache can assume that all SELECT statements, without exceptions, are cacheable.
  • verify_cacheable: The cache can not assume that all SELECT statements are cacheable, but must verify that.
select=assume_cacheable

Default is verify_cacheable. In this case, the SELECT statements will be parsed and only those that are safe for caching - e.g. do not call any non-cacheable functions or access any non-cacheable variables - will be subject to caching.

If assume_cacheable is specified, then all SELECT statements are assumed to be cacheable and will be parsed only if some specific rule requires that.

debug

An integer value, using which the level of debug logging made by the cache can be controlled. The value is actually a bitfield with different bits denoting different logging.

  • 0 (0b00000) No logging is made.
  • 1 (0b00001) A matching rule is logged.
  • 2 (0b00010) A non-matching rule is logged.
  • 4 (0b00100) A decision to use data from the cache is logged.
  • 8 (0b01000) A decision not to use data from the cache is logged.
  • 16 (0b10000) Higher level decisions are logged.

Default is 0. To log everything, give debug a value of 31.

debug=31

Rules

The caching rules are expressed as a JSON object.

There are two decisions to be made regarding the caching; in what circumstances should data be stored to the cache and in what circumstances should the data in the cache be used.

In the JSON object this is visible as follows:

{
    store: [ ... ],
    use: [ ... ]
}

The store field specifies in what circumstances data should be stored to the cache and the use field specifies in what circumstances the data in the cache should be used. In both cases, the value is a JSON array containg objects.

When to Store

By default, if no rules file have been provided or if the store field is missing from the object, the results of all queries will be stored to the cache, subject to max_resultset_rows and max_resultset_size cache filter parameters.

By providing a store field in the JSON object, the decision whether to store the result of a particular query to the cache can be controlled in a more detailed manner. The decision to cache the results of a query can depend upon

  • the database,
  • the table,
  • the column, or
  • the query itself.

Each entry in the store array is an object containing three fields,

{
    "attribute": <string>,
    "op": <string>
    "value": <string>
}

where, * the attribute can be database, table, column or query, * the op can be =, !=, like or unlike, and * the value a string.

If op is = or != then value is used as a string; if it is like or unlike, then value is interpreted as a pcre2 regular expression. Note though that if attribute is database, table or column, then the string is interpreted as a name, where a dot . denotes qualification or scoping.

The objects in the store array are processed in order. If the result of a comparison is true, no further processing will be made and the result of the query in question will be stored to the cache.

If the result of the comparison is false, then the next object is processed. The process continues until the array is exhausted. If there is no match, then the result of the query is not stored to the cache.

Note that as the query itself is used as the key, although the following queries

select * from db1.tbl

and

use db1;
select * from tbl

target the same table and produce the same results, they will be cached separately. The same holds for queries like

select * from tbl where a = 2 and b = 3;

and

select * from tbl where b = 3 and a = 2;

as well. Although they conceptually are identical, there will be two cache entries.

Qualified Names

When using = or != in the rule object in conjunction with database, table and column, the provided string is interpreted as a name, that is, dot (.) denotes qualification or scope.

In practice that means that if attribute is database then value may not contain a dot, if attribute is table then value may contain one dot, used for separating the database and table names respectively, and if attribute is column then value may contain one or two dots, used for separating table and column names, or database, table and column names.

Note that if a qualified name is used as a value, then all parts of the name must be available for a match. Currently Maria DB MaxScale may not always be capable of deducing in what table a particular column is. If that is the case, then a value like tbl.field may not necessarily be a match even if the field is field and the table actually is tbl.

Implication of the default database.

If the rules concerns the database, then only if the statement refers to no specific database, will the default database be considered.

Regexp Matching

The string used for matching the regular expression contains as much information as there is available. For instance, in a situation like

use somedb;
select fld from tbl;

the string matched against the regular expression will be somedb.tbl.fld.

Examples

Cache all queries targeting a particular database.

{
    "store": [
        {
            "attribute": "database",
            "op": "=",
            "value": "db1"
        }
    ]
}

Cache all queries not targeting a particular table

{
    "store": [
        {
            "attribute": "table",
            "op": "!=",
            "value": "tbl1"
        }
    ]
}

That will exclude queries targeting table tbl1 irrespective of which database it is in. To exclude a table in a particular database, specify the table name using a qualified name.

{
    "store": [
        {
            "attribute": "table",
            "op": "!=",
            "value": "db1.tbl1"
        }
    ]
}

Cache all queries containing a WHERE clause

{
    "store": [
        {
            "attribute": "query",
            "op": "like",
            "value": ".*WHERE.*"
        }
    ]
}

Note that that will actually cause all queries that contain WHERE anywhere, to be cached.

When to Use

By default, if no rules file have been provided or if the use field is missing from the object, all users may be returned data from the cache.

By providing a use field in the JSON object, the decision whether to use data from the cache can be controlled in a more detailed manner. The decision to use data from the cache can depend upon

  • the user.

Each entry in the use array is an object containing three fields,

{
    "attribute": <string>,
    "op": <string>
    "value": <string>
}

where, * the attribute can be user, * the op can be =, !=, like or unlike, and * the value a string.

If op is = or != then value is interpreted as a MariaDB account string, that is, % means indicates wildcard, but if op is like or unlike it is simply assumed value is a pcre2 regular expression.

For instance, the following are equivalent:

{
    "attribute": "user",
    "op": "=",
    "value": "'bob'@'%'"
}

{
    "attribute": "user",
    "op": "like",
    "value": "bob@.*"
}

Note that if op is = or != then the usual assumptions apply, that is, a value of bob is equivalent with 'bob'@'%'. If like or unlike is used, then no assumptions apply, but the string is used verbatim as a regular expression.

The objects in the use array are processed in order. If the result of a comparison is true, no further processing will be made and the data in the cache will be used, subject to the value of ttl.

If the result of the comparison is false, then the next object is processed. The process continues until the array is exhausted. If there is no match, then data in the cache will not be used.

Note that use is relevant only if the query is subject to caching, that is, if all queries are cached or if a query matches a particular rule in the store array.

Examples

Use data from the cache for all users except admin (actually 'admin'@'%'), regardless of what host the admin user comes from.

{
    "use": [
        {
            "attribute": "user",
            "op": "!=",
            "value": "admin"
        }
    ]
}

Security

As the cache is not aware of grants, unless the cache has been explicitly configured who the caching should apply to, the presence of the cache may provide a user with access to data he should not have access to.

Suppose there is a table access that the user alice has access to, but the user bob does not. If bob tries to access the table, he will get an error as reply:

MySQL [testdb]> select * from access;
ERROR 1142 (42000): SELECT command denied to user 'bob'@'localhost' for table 'access'

If we now setup caching for the table, using the simplest possible rules file, bob will get access to data from the table, provided he executes a select identical with one alice has executed.

For instance, suppose the rules look as follows:

{
    "store": [
        {
            "attribute": "table",
            "op": "=",
            "value": "access"
        }
    ]
}

If alice now queries the table, she will get the result, which also will be cached:

MySQL [testdb]> select * from access;
+------+------+
| a    | b    |
+------+------+
|   47 |   11 |
+------+------+

If bob now executes the very same query, and the result is still in the cache, it will be returned to him.

MySQL [testdb]> select current_user();
+----------------+
| current_user() |
+----------------+
| bob@127.0.0.1  |
+----------------+
1 row in set (0.00 sec)

MySQL [testdb]> select * from access;
+------+------+
| a    | b    |
+------+------+
|   47 |   11 |
+------+------+

That can be prevented, by explicitly declaring in the rules that the caching should be applied to alice only.

{
    "store": [
        {
            "attribute": "table",
            "op": "=",
            "value": "access"
        }
    ],
    "use": [
        {
            "attribute": "user",
            "op": "=",
            "value": "'alice'@'%'"
        }
    ]
}

With these rules in place, bob is again denied access, since queries targeting the table access will in his case not be served from the cache.

Storage

storage_inmemory

This simple storage module uses the standard memory allocator for storing the cached data.

storage=storage_inmemory

storage_rocksdb

This storage module uses RocksDB database for storing the cached data. The directory where the RocksDB database will be created is by default created into the MaxScale cache directory, which usually is not on a RAM disk. For maximum performance, you may want to explicitly place the RocksDB database on a RAM disk.

storage=storage_rocksdb

Parameters

cache_directory

Specifies the directory under which the filter instance specific RocksDB databases will be placed. Note that at startup, each RocksDB database will be deleted and recreated. That is, cache content will not be retained across MaxScale restarts.

storage_options=cache_directory=/mnt/maxscale-cache

With the above setting a directory /mnt/macscale-cache/storage_rocksdb will created, under which the actual instance specific cache directories are created.

collect_statistics

Specifies whether RocksDB should collect statistics that later can be queried using maxadmin. It should be noted, though, that collecting RocksDB statistics is not without a cost. From the RocksDB Documentation

The overhead of statistics is usually small but non-negligible. We usually observe an overhead of 5%-10%.

The value is a boolean and the default is false.

storage_options=collect_statistics=true

Example

In the following we define a cache MyCache that uses the cache storage module storage_inmemory and whose soft ttl is 30 seconds and whose hard ttl is 45 seconds. The cached data is shared between all threads and the maximum size of the cached data is 50 mebibytes. The rules for the cache are in the file cache_rules.json.

Configuration

[MyCache]
type=filter
module=cache
storage=storage_inmemory
soft_ttl=30
hard_ttl=45
cached_data=shared
max_size=50Mi
rules=cache_rules.json

[MyService]
type=service
...
filters=MyCache

cache_rules.json

The rules specify that the data of the table sbtest should be cached.

{
    "store": [
        {
            "attribute": "table",
            "op": "=",
            "value": "sbtest"
        }
    ]
}

Performance

Perhaps the most significant factor affecting the performance of the cache is whether the statements need to be parsed or not. By default, all statements are parsed in order to exclude SELECT statements that use non-cacheable functions, access non-cacheable variables or refer to system or user variables.

If it is known that no such statements are used or if it does not matter if the results are cached, that safety measure can be turned off. To do that, add the following line to the cache configuration:

[MyCache]
...
selects=assume_cacheable

With that configuration, the cache itself will not cause the statements to be parsed.

But note that even with assume_cacheable configured, a rule referring specifically to a database, table or column will still cause the statement to be parsed.

For instance, a simple rule like

{
    "store": [
        {
            "attribute": "database",
            "op": "=",
            "value": "db1"
        }
    ]
}

cannot be fulfilled without parsing the statement.

If the rule is instead expressed using a regular expression

{
    "store": [
        {
            "attribute": "query",
            "op": "like",
            "value": "FROM db1\\..*"
        }
    ]
}

then the statement will again not be parsed.

However, even if regular expression matching performance wise is cheaper than parsing, it still carries a cost. In the following is a table with numbers giving a rough picture of the relative cost of different approaches.

In the table, regexp match means that the cacheable statements were picked out using a rule like

{
    "attribute": "query",
    "op": "like",
    "value": "FROM dbname"
}

while exact match means that the cacheable statements were picked out using a rule like

{
    "attribute": "database",
    "op": "=",
    "value": "dbname"
}

The exact match rule requires all statements to be parsed.

Note that the qps figures are only indicative.

selects Rule qps
assume_cacheable none 100
assume_cacheable regexp match 98
assume_cacheable exact match 60
verify_cacheable none 60
verify_cacheable regexp match 58
verify_cacheable exact match 58

Summary

For maximum performance: Arrange the situation so that selects=assume_cacheable can be configured, and use no rules. If selects=assume_cacheable has been configured, use only regexp based rules. * If selects=verify_cacheable has been configured non-regex based matching can be used.

Comments

Comments loading...
Loading