Apache Kudu Configuration Reference

block cache capacity in MB

Whether the Log/WAL should explicitly call fsync() after each write.

Comma-separated list of directories with data blocks. If this is not specified, fs_wal_dir will be used as the sole data block directory.

Directory with write-ahead logs. If this is not specified, the program will not start. May be the same as fs_data_dirs

Comma-separated list of addresses to bind to for RPC connections. Currently, ephemeral ports (i.e. port 0) are not allowed.

Port to bind to for the web server

Prefix of log filename - full path is <log_dir>/<log_filename>.[INFO|WARN|ERROR|FATAL]

Size of the maintenance manager thread pool. For spinning disks, the number of threads should not be above the number of devices.

Maximum amount of memory this daemon should use, in bytes. A value of 0 autosizes based on the total system memory. A value of -1 disables all memory limiting.

load flags from file

show help on all flags [tip: all flags can have two dashes]

show version and build info and exit

color messages logged to stderr (if supported by terminal)

If specified, logfiles are written into this directory instead of the default logging directory.

log messages go to stderr instead of logfiles

approx. maximum log file size (in MB). A value of 0 will be silently overridden to 1.

Enables heap profiling for the lifetime of the process. Profile output will be stored in the directory specified by -heap_profile_path. Enabling this option will disable the on-demand/remote server profile handlers.

Output path to store heap profiles. If not set profiles are stored in /tmp/<process-name>.<pid>.<n>.heap.

set flags from the environment [use 'export FLAGS_flag1=value']

set flags from the environment if present

comma-separated list of flag names that it is okay to specify on the command line even if the program does not define a flag with that name. IMPORTANT: flags in this list that have arguments MUST use the flag=value format

show help on modules whose name contains the specified substr

show help on the modules named by this flag value

show help on all modules in the main package

show help on only the main module for this program

produce an xml version of help

Put additional links to the log files in this directory

Prepend the log prefix to the start of each log line

Messages logged at a lower level than this don’t actually get logged anywhere

log messages at or above this level are copied to stderr in addition to logfiles. This flag obsoletes --alsologtostderr.

Stop attempting to log to disk if the disk is full.

Symbolize the stack trace in the tombstone

Show all VLOG(m) messages for m ⇐ this. Overridable by --vmodule.

per-module verbose level. Argument is a comma-separated list of <module name>=<log level>. <module name> is a glob pattern, matched against the filename base (that is, name ignoring .cc/.h./-inl.h). <log level> overrides any value given by --v.

block cache capacity in MB

Whether the Log/WAL should explicitly call fsync() after each write.

Comma-separated list of directories with data blocks. If this is not specified, fs_wal_dir will be used as the sole data block directory.

Directory with write-ahead logs. If this is not specified, the program will not start. May be the same as fs_data_dirs

Comma-separated list of addresses to bind to for RPC connections. Currently, ephemeral ports (i.e. port 0) are not allowed.

Port to bind to for the web server

Comma separated addresses of the masters which the tablet server should connect to. The masters do not read this flag — configure the masters separately using 'rpc_bind_addresses'.

Prefix of log filename - full path is <log_dir>/<log_filename>.[INFO|WARN|ERROR|FATAL]

Size of the maintenance manager thread pool. For spinning disks, the number of threads should not be above the number of devices.

Maximum amount of memory this daemon should use, in bytes. A value of 0 autosizes based on the total system memory. A value of -1 disables all memory limiting.

load flags from file

show help on all flags [tip: all flags can have two dashes]

show version and build info and exit

color messages logged to stderr (if supported by terminal)

If specified, logfiles are written into this directory instead of the default logging directory.

log messages go to stderr instead of logfiles

approx. maximum log file size (in MB). A value of 0 will be silently overridden to 1.

Enables heap profiling for the lifetime of the process. Profile output will be stored in the directory specified by -heap_profile_path. Enabling this option will disable the on-demand/remote server profile handlers.

Output path to store heap profiles. If not set profiles are stored in /tmp/<process-name>.<pid>.<n>.heap.

set flags from the environment [use 'export FLAGS_flag1=value']

set flags from the environment if present

comma-separated list of flag names that it is okay to specify on the command line even if the program does not define a flag with that name. IMPORTANT: flags in this list that have arguments MUST use the flag=value format

show help on modules whose name contains the specified substr

show help on the modules named by this flag value

show help on all modules in the main package

show help on only the main module for this program

produce an xml version of help

Put additional links to the log files in this directory

Prepend the log prefix to the start of each log line

Messages logged at a lower level than this don’t actually get logged anywhere

log messages at or above this level are copied to stderr in addition to logfiles. This flag obsoletes --alsologtostderr.

Stop attempting to log to disk if the disk is full.

Symbolize the stack trace in the tombstone

Show all VLOG(m) messages for m ⇐ this. Overridable by --vmodule.

per-module verbose level. Argument is a comma-separated list of <module name>=<log level>. <module name> is a glob pattern, matched against the filename base (that is, name ignoring .cc/.h./-inl.h). <log level> overrides any value given by --v.

Which type of block cache to use for caching data. Valid choices are 'DRAM' or 'NVM'. DRAM, the default, caches data in regular memory. 'NVM' caches data in a memory-mapped file using the NVML library.

The default block size to use in cfiles

Default cfile block compression codec.

What to do to cfile blocks when writing is finished. Possible values are 'close', 'flush', or 'nothing'.

Whether to print the LLVM IR for generated functions

Whether to dump the disassembly of the machine code for generated functions.

Number of entries which may be stored in the code generation cache.

Whether to print time that each code generation request took.

The maximum per-tablet RPC batch size when updating peers.

Seconds that a leader is unable to successfully heartbeat to a follower after which the follower is considered to be failed and evicted from the config.

Maximum size of the group commit queue in bytes

The minimum number of seconds for which to keep log segments to keep at all times, regardless of what is required for durability. Logs may be still retained for a longer amount of time if they are necessary for correct restart. This should be set long enough such that a tablet server which has temporarily failed can be restarted within the given time period. If a server is down for longer than this amount of time, it is possible that its tablets will be re-replicated on other machines.

The minimum number of past log segments to keep at all times, regardless of what is required for durability. Must be at least 1.

Server-wide version of 'log_cache_size_limit_mb'. The total memory used for caching log entries across all tablets is kept under this threshold.

The total per-tablet size of consensus entries which may be kept in memory. The log cache attempts to keep all entries which have not yet been replicated to all followers in memory, but if the total size of those entries exceeds this limit within an individual tablet, the oldest will be evicted.

Whether the WAL segments preallocation should happen asynchronously

Whether the WAL should preallocate the entire segment before writing to it

The default segment size for log roll-overs, in MB

Whether to evict followers from the Raft config that have fallen too far behind the leader’s log to catch up normally or have been unreachable by the leader for longer than follower_unavailable_considered_failed_sec

Maximum time to sleep in between leader election retries, in addition to the regular timeout. When leader election fails the interval in between retries increases exponentially, up to this value.

Maximum heartbeat periods that the leader can fail to heartbeat in before we consider the leader to be failed. The total failure timeout in milliseconds is raft_heartbeat_interval_ms times leader_failure_max_missed_heartbeat_periods. The value passed to this flag may be fractional.

The mean failure-checking interval of the randomized failure monitor. If this is configured to -1 (the default), uses the value of 'raft_heartbeat_interval_ms'.

The standard deviation of the failure-checking interval of the randomized failure monitor. If this is configured to -1 (the default), this is set to half of the mean check interval.

The heartbeat interval for Raft replication. The leader produces heartbeats to followers at this interval. The followers expect a heartbeat at this interval and consider a leader to have failed if it misses several in a row.

Coalesce synchronization of data during CloseBlocks()

Which block manager to use for storage. Valid options are 'file' and 'log'.

Number of seconds we cache the full-disk status in the block manager. During this time, writes to the corresponding root path will not be attempted.

Ensure hole punching is supported by the underlying filesystem

Maximum size (soft) of a log container

Number of bytes to preallocate in a log container when creating new blocks. Set to 0 to disable preallocation

Whether the master should delete tablets reported by tablet servers for which there are no corresponding records in the master’s metadata. Use this option with care; it may cause permanent tablet data loss under specific (and rare) cases of master failures!

Default number of replicas for tables that do not have the num_replicas set.

Amount of time to give a newly-elected leader master to load the previous master’s metadata and become active. If this time is exceeded, the node crashes.

Timeout used for the master→TS async rpc calls.

The number of tablets per TS that can be requested for a new table.

Maximum time in milliseconds which clients may cache table locations. New range partitions may not be visible to existing client instances until after waiting for the ttl period.

Timeout used by the master when attempting to create tablet replicas during table creation.

After this amount of time, the master will stop attempting to contact a tablet server in order to perform operations such as deleting a tablet.

Timeout for retrieving master registration over RPC.

Comma-separated list of all the RPC addresses for Master config. This is used to configure the replicated Master process (currently considered experimental). NOTE: if not specified, configures a non-replicated Master.

The period of time that a Master can go without receiving a heartbeat from a tablet server before considering it unresponsive. Unresponsive servers are not selected when assigning replicas during table creation or re-replication.

Socket backlog parameter used when listening for RPC connections. This defines the maximum length to which the queue of pending TCP connections inbound to the RPC server may grow. If a connection request arrives when the queue is full, the client may receive an error. Higher values may help the server ride over bursts of new inbound connection requests.

If an RPC connection from a client is idle for this amount of time, the server will disconnect the client.

If enabled, dump traces of all RPC negotiations to the log

The maximum number of cycles for which an RPC callback should be allowed to run without emitting a warning. (Advanced debugging option)

Timeout for negotiating an RPC connection.

Maximum amount of time, in milliseconds, the server "remembers" a client for the purpose of caching its responses. After this period without hearing from it, the client is no longer remembered and the memory occupied by its responses is reclaimed. Retries of requests older than 'remember_clients_ttl_ms' are treated as new ones.

Maximum amount of time, in milliseconds, the server "remembers" a response to a specific request for a client. After this period has elapsed, the response may have been garbage collected and the client might get a response indicating the request is STALE.

If true, dump all RPC traces at INFO level

The maximum size of a message that any RPC that the server will accept. Must be at least 1MB.

The maximum number of bytes to display on the debug webserver’s log page

Maximum allowed clock synchronization error as reported by NTP before the server will abort.

Number of RPC acceptor threads for each bound address

Number of RPC worker threads to run

Default length of queue for incoming RPC requests

Number of libev reactor threads to start.

Interval (in milliseconds) at which the server will dump its metrics to a local log file. The log files are located in the same directory as specified by the -log_dir flag. If this is not a positive value, then metrics logging will be disabled.

The maximum length of a POST request that will be accepted by the embedded web server.

Files under <webserver_doc_root> are accessible via the debug webserver. Defaults to $KUDU_HOME/www, or if $KUDU_HOME is not set, disables the document root

If true, webserver may serve static files from the webserver_doc_root

Interface to start debug webserver on. If blank, webserver binds to 0.0.0.0

Maximum number of threads to start for handling web server requests

The target size for DiskRowSets during flush/compact when the budgeted compaction policy is used

Block size for delta files. In the future, this may become configurable on a per-table basis.

Block size used for composite key indexes.

Minimum ratio of sizeof(deltas) to sizeof(base data) before a major compaction can run (Advanced option)

How many delta stores are required before forcing a minor delta compaction (Advanced option)

Block size of the bloom filters used for tablet keys.

Target false-positive rate (between 0 and 1) to size tablet key bloom filters. A lower false positive rate may reduce the number of disk seeks required in heavy insert workloads, at the expense of more space and RAM required for bloom filters.

Budget for a single compaction

Burst factor for write RPC throttling. The maximum rate the throttler allows within a token refill period (100ms) equals burst factor multiply base rate.

Maximum write RPC IO rate (byte/s) allowed for a tablet, write RPC exceeding this limit will be throttled. 0 means no limit.

Maximum write RPC rate (op/s) allowed for a tablet, write RPC exceeding this limit will be throttled. 0 means no limit.

Size at which MemRowSet flushes are triggered. A MRS can still flush below this threshold if it if hasn’t flushed in a while, or if the server-wide memory limit has been reached.

Maximum amount of memory that may be consumed by all in-flight transactions belonging to a particular tablet. When this limit is reached, new transactions will be rejected and clients will be forced to retry them. If -1, transaction memory tracking is disabled.

Interval at which the TS heartbeats to the master.

Maximum number of consecutive heartbeat failures until the Tablet Server backs off to the normal heartbeat interval, rather than retrying.

Timeout used for the TS→Master heartbeat RPCs.

Number of milliseconds of inactivity allowed for a scannerbefore it may be expired

The number of rows to batch for servicing scan requests.

The default size for batches of scan results

The maximum batch size that a client may request for scan results.

Number of threads available to open tablets during startup. If this is set to 0 (the default), then the number of bootstrap threads will be set based on the number of data directories. If the data directories are on some very fast storage device such as SSD or a RAID array, it may make sense to manually tune this.

Trace pattern specifying which trace events should be dumped directly to the console

Never fsync() anything to disk. This is used by certain test cases to speed up runtime. This is very unsafe to use in production.

Kill the process if an I/O operation results in EIO

Use fsync(2) instead of fdatasync(2) for synchronizing dirty data to disk.

Disable core dumps when this process crashes.

Percentage of the hard memory limit that this daemon may consume before memory throttling of writes begins. The greater the excess, the higher the chance of throttling. In general, a lower soft limit leads to smoother write latencies but decreased throughput, and vice versa for a higher soft limit.

Percentage of the hard memory limit that this daemon may consume before WARNING level messages are periodically logged.

Maximum percentage of the RSS that tcmalloc is allowed to use for reserved but unallocated memory.

The minimum number of milliseconds a metric will be kept for after it is no longer active. (Advanced option)

IP to bind to when making outgoing socket connections. This must be an IP address of the form A.B.C.D, not a hostname. Advanced parameter, subject to change.

The number of times that the NVM cache will retry attempts to allocate memory for new entries. In between attempts, a cache entry will be evicted.

The path at which the NVM cache will try to allocate its memory. This can be a tmpfs or ramfs for testing purposes.

Drop in-memory buffers of log contents. Logs can grow very quickly and they are rarely read before they need to be evicted from memory. Instead, drop them from memory as soon as they are flushed to disk.

Emit a backtrace when logging at file:linenum.

Buffer log messages logged at this level or lower (-1 means don’t buffer; 0 means buffer INFO only; …​)

Buffer log messages for at most this many seconds

Which type of block cache to use for caching data. Valid choices are 'DRAM' or 'NVM'. DRAM, the default, caches data in regular memory. 'NVM' caches data in a memory-mapped file using the NVML library.

The default block size to use in cfiles

Default cfile block compression codec.

What to do to cfile blocks when writing is finished. Possible values are 'close', 'flush', or 'nothing'.

Whether to print the LLVM IR for generated functions

Whether to dump the disassembly of the machine code for generated functions.

Number of entries which may be stored in the code generation cache.

Whether to print time that each code generation request took.

The maximum per-tablet RPC batch size when updating peers.

Seconds that a leader is unable to successfully heartbeat to a follower after which the follower is considered to be failed and evicted from the config.

Maximum size of the group commit queue in bytes

The minimum number of seconds for which to keep log segments to keep at all times, regardless of what is required for durability. Logs may be still retained for a longer amount of time if they are necessary for correct restart. This should be set long enough such that a tablet server which has temporarily failed can be restarted within the given time period. If a server is down for longer than this amount of time, it is possible that its tablets will be re-replicated on other machines.

The minimum number of past log segments to keep at all times, regardless of what is required for durability. Must be at least 1.

Server-wide version of 'log_cache_size_limit_mb'. The total memory used for caching log entries across all tablets is kept under this threshold.

The total per-tablet size of consensus entries which may be kept in memory. The log cache attempts to keep all entries which have not yet been replicated to all followers in memory, but if the total size of those entries exceeds this limit within an individual tablet, the oldest will be evicted.

Whether the WAL segments preallocation should happen asynchronously

Whether the WAL should preallocate the entire segment before writing to it

The default segment size for log roll-overs, in MB

Whether to evict followers from the Raft config that have fallen too far behind the leader’s log to catch up normally or have been unreachable by the leader for longer than follower_unavailable_considered_failed_sec

Maximum time to sleep in between leader election retries, in addition to the regular timeout. When leader election fails the interval in between retries increases exponentially, up to this value.

Maximum heartbeat periods that the leader can fail to heartbeat in before we consider the leader to be failed. The total failure timeout in milliseconds is raft_heartbeat_interval_ms times leader_failure_max_missed_heartbeat_periods. The value passed to this flag may be fractional.

The mean failure-checking interval of the randomized failure monitor. If this is configured to -1 (the default), uses the value of 'raft_heartbeat_interval_ms'.

The standard deviation of the failure-checking interval of the randomized failure monitor. If this is configured to -1 (the default), this is set to half of the mean check interval.

The heartbeat interval for Raft replication. The leader produces heartbeats to followers at this interval. The followers expect a heartbeat at this interval and consider a leader to have failed if it misses several in a row.

Coalesce synchronization of data during CloseBlocks()

Which block manager to use for storage. Valid options are 'file' and 'log'.

Number of seconds we cache the full-disk status in the block manager. During this time, writes to the corresponding root path will not be attempted.

Ensure hole punching is supported by the underlying filesystem

Maximum size (soft) of a log container

Number of bytes to preallocate in a log container when creating new blocks. Set to 0 to disable preallocation

Socket backlog parameter used when listening for RPC connections. This defines the maximum length to which the queue of pending TCP connections inbound to the RPC server may grow. If a connection request arrives when the queue is full, the client may receive an error. Higher values may help the server ride over bursts of new inbound connection requests.

If an RPC connection from a client is idle for this amount of time, the server will disconnect the client.

If enabled, dump traces of all RPC negotiations to the log

The maximum number of cycles for which an RPC callback should be allowed to run without emitting a warning. (Advanced debugging option)

Timeout for negotiating an RPC connection.

Maximum amount of time, in milliseconds, the server "remembers" a client for the purpose of caching its responses. After this period without hearing from it, the client is no longer remembered and the memory occupied by its responses is reclaimed. Retries of requests older than 'remember_clients_ttl_ms' are treated as new ones.

Maximum amount of time, in milliseconds, the server "remembers" a response to a specific request for a client. After this period has elapsed, the response may have been garbage collected and the client might get a response indicating the request is STALE.

If true, dump all RPC traces at INFO level

The maximum size of a message that any RPC that the server will accept. Must be at least 1MB.

The maximum number of bytes to display on the debug webserver’s log page

Maximum allowed clock synchronization error as reported by NTP before the server will abort.

Number of RPC acceptor threads for each bound address

Number of RPC worker threads to run

Default length of queue for incoming RPC requests

Number of libev reactor threads to start.

Interval (in milliseconds) at which the server will dump its metrics to a local log file. The log files are located in the same directory as specified by the -log_dir flag. If this is not a positive value, then metrics logging will be disabled.

The maximum length of a POST request that will be accepted by the embedded web server.

Files under <webserver_doc_root> are accessible via the debug webserver. Defaults to $KUDU_HOME/www, or if $KUDU_HOME is not set, disables the document root

If true, webserver may serve static files from the webserver_doc_root

Interface to start debug webserver on. If blank, webserver binds to 0.0.0.0

Maximum number of threads to start for handling web server requests

The target size for DiskRowSets during flush/compact when the budgeted compaction policy is used

Block size for delta files. In the future, this may become configurable on a per-table basis.

Block size used for composite key indexes.

Minimum ratio of sizeof(deltas) to sizeof(base data) before a major compaction can run (Advanced option)

How many delta stores are required before forcing a minor delta compaction (Advanced option)

Block size of the bloom filters used for tablet keys.

Target false-positive rate (between 0 and 1) to size tablet key bloom filters. A lower false positive rate may reduce the number of disk seeks required in heavy insert workloads, at the expense of more space and RAM required for bloom filters.

Budget for a single compaction

Burst factor for write RPC throttling. The maximum rate the throttler allows within a token refill period (100ms) equals burst factor multiply base rate.

Maximum write RPC IO rate (byte/s) allowed for a tablet, write RPC exceeding this limit will be throttled. 0 means no limit.

Maximum write RPC rate (op/s) allowed for a tablet, write RPC exceeding this limit will be throttled. 0 means no limit.

Size at which MemRowSet flushes are triggered. A MRS can still flush below this threshold if it if hasn’t flushed in a while, or if the server-wide memory limit has been reached.

Maximum amount of memory that may be consumed by all in-flight transactions belonging to a particular tablet. When this limit is reached, new transactions will be rejected and clients will be forced to retry them. If -1, transaction memory tracking is disabled.

Interval at which the TS heartbeats to the master.

Maximum number of consecutive heartbeat failures until the Tablet Server backs off to the normal heartbeat interval, rather than retrying.

Timeout used for the TS→Master heartbeat RPCs.

Number of milliseconds of inactivity allowed for a scannerbefore it may be expired

The number of rows to batch for servicing scan requests.

The default size for batches of scan results

The maximum batch size that a client may request for scan results.

Number of threads available to open tablets during startup. If this is set to 0 (the default), then the number of bootstrap threads will be set based on the number of data directories. If the data directories are on some very fast storage device such as SSD or a RAID array, it may make sense to manually tune this.

Trace pattern specifying which trace events should be dumped directly to the console

Never fsync() anything to disk. This is used by certain test cases to speed up runtime. This is very unsafe to use in production.

Kill the process if an I/O operation results in EIO

Use fsync(2) instead of fdatasync(2) for synchronizing dirty data to disk.

Disable core dumps when this process crashes.

Percentage of the hard memory limit that this daemon may consume before memory throttling of writes begins. The greater the excess, the higher the chance of throttling. In general, a lower soft limit leads to smoother write latencies but decreased throughput, and vice versa for a higher soft limit.

Percentage of the hard memory limit that this daemon may consume before WARNING level messages are periodically logged.

Maximum percentage of the RSS that tcmalloc is allowed to use for reserved but unallocated memory.

The minimum number of milliseconds a metric will be kept for after it is no longer active. (Advanced option)

IP to bind to when making outgoing socket connections. This must be an IP address of the form A.B.C.D, not a hostname. Advanced parameter, subject to change.

The number of times that the NVM cache will retry attempts to allocate memory for new entries. In between attempts, a cache entry will be evicted.

The path at which the NVM cache will try to allocate its memory. This can be a tmpfs or ramfs for testing purposes.

Drop in-memory buffers of log contents. Logs can grow very quickly and they are rarely read before they need to be evicted from memory. Instead, drop them from memory as soon as they are flushed to disk.

Emit a backtrace when logging at file:linenum.

Buffer log messages logged at this level or lower (-1 means don’t buffer; 0 means buffer INFO only; …​)

Buffer log messages for at most this many seconds