Using the S3 Storage Engine
- New Options for CREATE TABLE
- mysqld Startup Options for S3
- Typical my.cnf Entry
- Typical Usage Case for S3 Tables
- Operations Allowed on S3 Tables
- Current Limitations
- Performance Considerations
- Future Development Ideas
- Troubleshooting S3 on SELinux
- See Also
The S3 storage engine is read only and allows one to archive MariaDB tables in Amazon S3, or any third-party public or private cloud that implements S3 API (of which there are many), but still have them accessible for reading in MariaDB.
To move data to S3 one can do
ALTER TABLE old_table ENGINE=S3
To get data back to a 'normal' table one can do:
ALTER TABLE s3_table ENGINE=INNODB
New Options for CREATE TABLE
S3_BLOCK_SIZE: Set to 4M as default. This is the block size for all index and data pages stored in S3.
COMPRESSION_ALGORITHM: Set to 'none' as default. Which compression algorithm to use for block stored in S3. Options are:
mysqld Startup Options for S3
To be able to use S3 for storage one *must* define how to access S3 and where data are stored in S3:
- s3_access_key: The AWS access key to access your data
- s3_secret_key: The AWS secret key to access your data
- s3_bucket: The AWS bucket where your data should be stored. All MariaDB table data is stored in this bucket.
- s3_region: The AWS region where your data should be stored.
Other, less critical options, are:
- s3_host_name: Hostname for the S3 service. "s3.amazonaws.com", Amazon S3 service, by default.
- s3_protocol_version: Protocol used to communication with S3. One of "Auto", "Amazon" or "Original" where "Auto" is the default. If you get errors like "8 Access Denied" when you are connecting to another service provider, then try to change this option. The reason for this variable is that Amazon has changed some parts of the S3 protocol since they originally introduced it but other service providers are still using the original protocol.
- s3_block_size: Set to 4M as default. This is the default block size for a table, if not specified in CREATE TABLE.
- s3_pagecache_buffer_size: Default 128M. The size of the buffer used for index blocks for S3 tables. Increase this to get better index handling (for all reads and multiple writes) to as much as you can afford.
Last some options you probably don't have to ever touch:
- s3_pagecache_age_threshold : Default 300: This characterizes the number of hits a hot block has to be untouched until it is considered aged enough to be downgraded to a warm block. This specifies the percentage ratio of that number of hits to the total number of blocks in the page cache.
- s3_pagecache_division_limit: Default 100. The minimum percentage of warm blocks in key cache.
- s3_pagecache_file_hash_size: Default 512. Number of hash buckets for open files. If you have a lot of S3 files open you should increase this for faster flush of changes. A good value is probably 1/10 of number of possible open S3 files.
- s3_debug: Default 0. Generates a trace file from libmarias3 on stderr (mysqld.err) for debugging the S3 protocol.
Typical my.cnf Entry
[mysqld-10.5] s3=ON s3-bucket=mariadb s3-access-key=xxxx s3-secret-key=xxx s3-region=eu-north-1 #s3-host-name=s3.amazonaws.com #s3-protocol-version=Auto [aria_s3_copy] s3-bucket=mariadb s3-access-key=xxxx s3-secret-key=xxx s3-region=eu-north-1 #s3-host-name=s3.amazonaws.com #s3-protocol-version=Auto
Typical Usage Case for S3 Tables
The typical use case would be that there exists tables that after some time would become fairly inactive, but are still important so that they can not be removed. In that case, an option is to move such a table to an archiving service, which is accessible through an S3 API.
Notice that S3 means the Cloud Object Storage API defined by Amazon AWS. Often the whole of Amazon’s Cloud Object Storage is referred to as S3. In the context of the S3 archive storage engine, it refers to the API itself that defines how to store objects in a cloud service, being it Amazon’s or someone else’s. OpenStack for example provides an S3 API for storing objects.
The main benefit of storing things in an S3 compatible storage is that the cost of storage is much cheaper than many other alternatives. Many S3 implementations also provide reliable long-term storage.
Operations Allowed on S3 Tables
- ALTER TABLE S3 supports all types, keys and other options that are supported by the Aria engine.
- DROP TABLE
- SELECT Any SELECT operations you can perform on a normal table should work with an S3 table.
- SHOW TABLES will show all tables that exist in the current defined S3 location.
- S3 tables can be part of partitions.
The S3 storage engine supports full MariaDB discovery. This means that if you have the S3 storage engine enabled and properly configured, the table stored in S3 will automatically be discovered when it's accessed with SHOW TABLES, SELECT or any other operation that tries to access it. In the case of SELECT, the .frm file from S3 will be copied to the local storage to speed up future accesses.
When an S3 table is opened for the first time (it's not in the table cache) and there is a local .frm file, the S3 engine will check if it's still relevant, and if not, update or delete the .frm file.
This means that if the table definition changes on S3 and it's in the
local cache, one has to execute
FLUSH TABLES to
get MariaDB to notice the change and update the .frm file.
Discovery of S3 tables is not done for tables in the mysql databases to make mysqld boot faster and more securely.
- mysqldump will by default ignore S3 tables. If
mysqldumpis run with the
--copy-s3-tablesoption, the resulting file will contain a CREATE statement for a similar Aria table, followed by the table data and ending with an
ALTER TABLE xxx ENGINE=S3.
- One needs to do
FLUSH TABLESfor the MariaDB to notice that the definition has changed if the S3 table is in the table cache.
- Replication of S3 engines to a slave that uses the same S3 storage doesn't work. The slave will stop at any
ALTER TABLE ... ENGINE=S3with an error that the table already exists in S3.
- mysql-test-run doesn't by default test the S3 engine as we can't embed AWS keys into mysql-test-run.
Depending on your connection speed to your S3 provider, there can be some notable slowdowns in some operations.
As S3 is supporting discovery (automatically making tables available that are in S3) this can cause some small performance problems if the S3 engine is enabled.
- CREATE TABLE is a bit slower as the S3 engine has to check if the to-be-created table is already S3.
- Queries on information_schema tables are slower as S3 has to check if there is new tables in S3.
- DROP of non existing tables are slower as S3 has to check if the table is in S3.
There are no performance degradations when accessing existing tables on the server. Accessing the S3 table the first time will copy the .frm file from S3 to the local disk, speeding up future accesses to the table.
- Accessing a table on S3 can take some time , especially if you are using big packets (s3_block_size). However the second access to the same data should be fast as it's then cached in the S3 page cache.
Things to Try to Increase Performance
If you have performance problems with the S3 engine, here are some things you can try:
- Decreasing s3_block_size. This can be done both globally and per table.
- Use COMPRESSION_ALGORITHM=zlib when creating the table. This will decrease the amount of data transferred from S3 to the local cache.
- Increasing the size of the s3 page cache: s3_pagecache_buffer_size
Try also to execute the query twice to check if the problem is that the data was not properly cached. When data is cached locally the performance should be excellent.
Future Development Ideas
- Add MERGE support for S3 tables (possibly with another storage engine name).
- Store aws keys and region in the mysql.servers table (as Spider and FederatedX). This will allow one to have different tables on different S3 servers.
- Make s3 bucket, access_key and secret key thread specific
Troubleshooting S3 on SELinux
If you get errors such as:
ERROR 3 (HY000): Got error from put_object(bubu/produkt/frm): 5 Couldn't connect to server
one reason could be that your system doesn't allow MariaDB to connect to ports other than 3306.
The fix is to change