Backend Configuration

LevelDB

Configurable parameters for Riak’s LevelDB storage backend.

Note on upgrading to 2.0

If you are upgrading to Riak 2.0+ from a 1.x version, using LevelDB, and wish to use your old configuration files, i.e. app.config and vm.args, please note that you must set the total_leveldb_mem_percent setting in the eleveldb section of app.config. We recommend setting it to 70. If you do not set this parameter, it will default to 15, which can lead to problems in some clusters.

Config Description Default
leveldb.block_cache_threshold This setting defines the limit past which block cache memory can no longer be released in favor of the page cache. This setting has no impact in favor of file cache. The value is set on a per-vnode basis. 32MB
leveldb.compaction.trigger.tombstone_count Controls when a background compaction initiates solely due to the number of delete tombstones within an individual .sst table file. A value of off disables the feature. 1000
leveldb.compression Enabling this setting (on), which is the default, saves disk space. Disabling it may reduce read latency but increase overall disk activity. This option can be changed at any time, but it will not impact data on disk until the next time a file requires compaction. on
leveldb.compression.algorithm This setting is used to select which compression algorithm is selected when leveldb.compression is on. In new riak.conf files, this is explicitly set to lz4; however when this setting is not provided, snappy will be used for backward-compatibility.

When you determine that you will no longer need backward-compatibility, setting this to lz4 will cause future compactions to use the LZ4 algorithm for compression.
lz4 in new riak.conf files

snappy when not provided
leveldb.data_root The directory in which LevelDB will store its data. ./data/leveldb
leveldb.fadvise_willneed Option to override LevelDB's use of fadvise(DONTNEED) with fadvise(WILLNEED) instead. WILLNEED can reduce disk activity on systems where physical memory exceeds the database size. false
leveldb.maximum_memory This parameter defines the server memory (in bytes) to assign to LevelDB. Also see leveldb.maximum_memory.percent to set LevelDB memory as a percentage of system total. 80
leveldb.maximum_memory.percent This parameter defines the percentage of total server memory to assign to LevelDB. LevelDB will dynamically adjust its internal cache sizes to stay within this size. The memory size can alternately be assigned as a byte count via leveldb.maximum_memory instead. 70
leveldb.threads The number of worker threads performing LevelDB operations. 71
leveldb.verify_checksums Enables or disables the verification of the data fetched from LevelDB against internal checksums. on
leveldb.verify_compaction Enables or disables the verification of LevelDB data during compaction. on
leveldb.block.size_steps Defines the number of incremental adjustments to attempt between the block.size value and the maximum block.size for an .sst table file. A value of zero disables the underlying dynamic block_size feature. 16
leveldb.block.restart_interval Defines the key count threshold for a new key entry in the key index for a block. Most deployments should leave this parameter alone. 16
leveldb.block.size Defines the size threshold for a block/chunk of data within one .sst table file. Each new block gets an index entry in the .sst table file's master index. 4KB
leveldb.bloomfilter Each database .sst table file can include an optional "bloom filter" that is highly effective in shortcutting data queries that are destined to not find the requested key. The Bloom filter typically increases the size of an .sst table file by about 2%. on
leveldb.write_buffer_size_min Each vnode first stores new key/value data in a memory-based write buffer. This write buffer is in parallel to the recovery log mentioned in the sync parameter. Riak creates each vnode with a randomly sized write buffer for performance reasons. The random size is somewhere between write_buffer_size_min and write_buffer_size_max. 30MB
leveldb.write_buffer_size_max See leveldb.write_buffer_size_min directly above. 60MB
leveldb.limited_developer_mem This is a Riak-specific option that is used when a developer is testing a high number of vnodes and/or several VMs on a machine with limited physical memory. Do not use this option if making performance measurements. This option overwrites values given to write_buffer_size_min and write_buffer_size_max. off
leveldb.sync_on_write Whether LevelDB will flush after every write.

Note: If you are familiar with fsync, this is analogous to calling fsync after every write.
off
leveldb.tiered The level number at which LevelDB data switches from the faster to the slower array. The default of off disables the feature. off
leveldb.tiered.path.fast The path prefix for .sst files below the level set by leveldb.tiered.
leveldb.tiered.path.slow The path prefix for .sst files below the level set by leveldb.tiered.

Leveled

Configurable Parameters for Riak’s leveled storage backend

## Bitcask Configurable parameters for Riak's [Bitcask][plan backend bitcask] storage backend.
Config Description Default
leveled.data_root A path under which leveled data files will be stored. $(platform_data_dir)/leveled
leveled.sync_strategy Strategy for flushing data to disk - Can be set to riak_sync, sync (if OTP > 16) or none. Use none, and the OS will flush when most efficient. Use riak_sync or sync to flush after every PUT (not recommended wihtout some hardware support e.g. flash drives and/or Flash-backed Write Caches) none
leveled.compression_method Can be lz4 or native (which will use the Erlang native zlib compression) within term_to_binary native
leveled.compression_point The point at which compression is applied to the Journal (the Ledger is always compressed). Use on_receipt or on_compact. on_compact is suitable when values are unlikely to yield much benefit from compression(compression is only attempted when compacting) on_receipt
leveled.log_level Can be debug, info, warn, error or critical. Set the minimum log level to be used within leveled. Leveled will log many lines to allow for stats to be etracted by those using log indexers such as Splunk info
leveled.journal_size The approximate size (in bytes) when a Journal file should be rolled. Normally keep this as around the size of o(100K) objects. 1000000000
leveled.compaction_runs_perday The number of journal compactions per vnode per day, The higher the value, the more compaction runs, and the sooner space is recovered. But each run has a cost. 24
leveled.compaction_low_hour The hour of the day in which journal compaction can start. Use Low hour of 0 and High hour of 23 to have no compaction window (i.e. always compactregardless of time of day) 0
leveled.compaction_top_hour The hour of the day, after which journal compaction should stop. If low hour > top hour then, compaction will work overnight between low hour and top hour (inclusive). Timings rely on server's view of local time 23
leveled.max_run_length In a single compaction run, what is the maximum number of consecutive files which may be compacted. 4
Config Description Default
bitcask.data_root The directory under which Bitcask will store its data. ./data/bitcask
bitcask.io_mode Configure how Bitcask writes data to disk. If set to erlang, writes are made via Erlang's built-in file API; if set to nif, writes are made via direct calls to the POSIX C API. The nif mode provides higher throughput for certain workloads, but has the potential to negatively impact the Erlang VM, leading to higher worst-case latencies and possible throughput collapse erlang
bitcask.expiry By default, Bitcask keeps all of your data around. If your data has limited time value, or if you need to purge data for space reasons, you can set the expiry option. For example, if you need to purge data automatically after 1 day, set the value to 1d. off disables automatic expiration off
bitcask.expiry.grace_time By default, Bitcask will trigger a merge whenever a data file contains an expired key. This may result in excessive merging under some usage patterns. To prevent this you can set the bitcask.expiry.grace_time option. Bitcask will defer triggering a merge solely for key expiry by the configured number of seconds. Setting this to 1h effectively limits each cask to merging for expiry once per hour. 0
bitcask.hintfile_checksums Whether to allow the CRC to be present at the end of hintfiles. Setting this to allow_missing runs Bitcask in a backwards-compatible mode in which old hint files will still be accepted without CRC signatures. strict
bitcask.fold.max_puts See the description for the bitcask.fold.max_age config directly below. 0
bitcask.fold.max_age Fold keys thresholds will reuse the keydir if another fold was started less than fold.max_age ago and there were fewer than fold.max_puts updates. Otherwise, it will wait until all current fold keys complete and then start. Set either option to unlimited to disable. unlimited
bitcask.merge.thresholds.fragmentation Describes which ratio of dead keys to total keys in a file will cause it to be included in the merge. The value of this setting is a percentage from 0 to 100. For example, if a data file contains 4 dead keys and 6 live keys, it will be included in the merge at the default ratio (which is 40). Increasing the value will cause fewer files to be merged, decreasing the value will cause more files to be merged. 40
bitcask.merge.thresholds.dead_bytes Describes the minimum amount of data occupied by dead keys in a file to cause it to be included in the merge. Increasing the value will cause fewer files to be merged, whereas decreasing the value will cause more files to be merged. 128MB
bitcask.merge.thresholds.small_file Describes the minimum size a file must have to be excluded from the merge. Files smaller than the threshold will be included. Increasing the value will cause more files to be merged, whereas decreasing the value will cause fewer files to be merged. 10MB
bitcask.merge.triggers.dead_bytes Describes how much data stored for dead keys in a single file will trigger merging. If a file meets or exceeds the trigger value for dead bytes, merge will be triggered. Increasing the value will cause merging to occur less often, whereas decreasing the value will cause merging to happen more often. When either of these constraints are met by any file in the directory, Bitcask will attempt to merge files. 512MB
bitcask.merge.triggers.fragmentation Describes which ratio of dead keys to total keys in a file will trigger merging. The value of this setting is a percentage from 0 to 100. For example, if a data file contains 6 dead keys and 4 live keys, then merge will be triggered at the default setting. Increasing this value will cause merging to occur less often, whereas decreasing the value will cause merging to happen more often. 60
bitcask.merge.window.end See the description of the bitcask.merge.policy config below. 23
bitcask.merge.window.start See the description of the bitcask.merge.policy config below. 0
bitcask.merge.policy Lets you specify when during the day merge operations are allowed to be triggered. Valid options are: always, meaning no restrictions; never, meaning that merging will never be attempted; and window, specifying the hours during which merging is permitted, where bitcask.merge.window.start and bitcask.merge.window.end are integers between 0 and 23. If merging has a significant impact on performance of your cluster, or your cluster has quiet periods in which little storage activity occurs, you may want to change this setting from the default. always
bitcask.merge_check_interval Bitcask periodically runs checks to determine whether merges are necessary. This parameter determines how often those checks take place. Expressed as a time unit, e.g. `10s` for 10 seconds, `5m` for 5 minutes, etc. 3m
bitcask.merge_check_jitter In order to prevent merge operations from taking place on different nodes at the same time, Riak can apply random variance to merge times, expressed as a percentage of bitcask.merge_check_interval. 30%
bitcask.max_merge_size Maximum amount of data to merge in one go in the Bitcask backend. 100GB
bitcask.max_file_size Describes the maximum permitted size for any single data file in the Bitcask directory. If a write causes the current file to exceed this size threshold then that file is closed, and a new file is opened for writes. 2GB
bitcask.sync.interval See the description of the bitcask.sync.strategy directly below.
bitcask.sync.strategy Changes the durability of writes by specifying when to synchronize data to disk. The default setting protects against data loss in the event of application failure (process death) but leaves open a small window in which data could be lost in the event of complete system failure (e.g. hardware, OS, or power). The default mode, none, writes data into operating system buffers which will be written to the disks when those buffers are flushed by the operating system. If the system fails, e.g. due power loss or crash, that data is lost before those buffers are flushed to stable storage. This is prevented by the setting o_sync, which forces the operating system to flush to stable storage at every write. The effect of flushing each write is better durability, however write throughput will suffer as each write will have to wait for the write to complete. Available sync strategies: none, which will let the operating system manage syncing writes; o_sync, which will uses the O_SYNC flag to force syncs on every write; and interval, by which will force Bitcask to sync every bitcask.sync.interval seconds. none
bitcask.open_timeout Specifies the maximum time Bitcask will block on startup while attempting to create or open the data directory. You generally need not change this value. If for some reason the timeout is exceeded on open you'll see a log message of the form Failed to start bitcask backend: .... . Only then should you consider a longer timeout. 4s

Memory Backend

Configurable parameters for Riak’s Memory backend.

Config Description Default
memory_backend.ttl Each value written will be written with this "time to live." Once that object's time is up, it will be deleted on the next read of its key. Minimum: 1s.
memory_backend.max_memory_per_vnode The maximum amount of memory consumed per vnode by the memory storage backend. Minimum: 1MB.

Multi Backend

Configurable parameters for Riak’s Multi backend, which enables you to utilize multiple data backends in a single Riak cluster.

If you are using multiple backends, you can configure the backends individually by prepending the setting with multi_backend.$name, where $name is the name of the backend. $name can be any valid configuration word, like customer_data, my_data, foo_bar_backend, etc.

Below is the general form for setting multi-backend parameters:

multi_backend.$name.(existing_setting) = <setting>
# or
multi_backend.$name.$backend_type.(backend_specific_setting) = <setting>

Below is a listing of the available parameters:

Config Description Default
multi_backend.$name.storage_backend This parameter specifies the Erlang module defining the storage mechanism that will be used on this node. bitcask
multi_backend.default The default name of a backend when one is not specified.

To give an example, if you have a LevelDB backend named customer_backend and wish to set the data_root parameter to $(platform_data_dir)/leveldb_backends/customer_backend/, you would do so as follows:

multi_backend.customer_backend.storage_backend = leveldb
multi_backend.customer_backend.leveldb.data_root = $(platform_data_dir)/leveldb_backends/customer_backend
multi_backend.customer_backend.leveldb.maximum_memory.percent = 50