«

Monitoring and Performance

This page covers

Metric names are identified in the text by surrounding square brackets, e.g. [TransactionMsec]. Each reported value of a metric can include up to five different statistics, which are capitalized where they appear in the text: Samples, Average, Minimum, Maximum, and Sum.

monitoring.svg

Tuning Tips

Give your peers and transactor enough memory

You can set the maximum memory available to a JVM process with the -Xmx flag to java (or to bin/transactor). The more memory you give a process, the more database index it will be able to cache locally in memory, avoiding trips to memcached or storage.

Use memcached

Memcached is a great fit for Datomic. Read more about configuring memcached in the Caching documentation.

Give memcached enough memory

Memcached can deliver data to peers with very low latency (on the order of 1 millisecond for a segment). If you can fit your entire database (or even the commonly-used working set) in memcached, your read and query performance will be quite good.

Diagnostic: If your [Memcache] Average is substantially less than 1.0, then you can serve more reads from memcache by increasing the size of your memcached cluster.

Add more peers

Datomic read and query is horizontally scalable. If you can't serve as many simultaneous queries as you like, add more peers.

Add more cores

Datomic's entire public API is threadsafe, and is designed to automatically take advantage of multiple cores where available. Having more cores will help any Datomic process, transactor or peer.

Don't benchmark against the memory database

The memory database is a convenience for development, testing, and temp data. Memory is not good for storage-sized datasets, as memory isn't storage-sized.

Provision enough DynamoDB capacity

See the DynamoDB section of Capacity Planning.

Run large imports on cheap storage, then backup/restore

Restoring a database consumes significantly less write capacity than building the database via transactions. So if your target system has metered storage pricing (e.g. DynamoDB), it may make sense to do large import jobs on less expensive storage, backup, and then restore to the target storage.

Use metrics to analyze write activity

The speed with which a Datomic transactor can write data is measured with the [TransactionMsec] metric. The minimum, average, and maximum values of this metric, combined with the size of your transactions, characterize the write latency of the system from the transactor's perspective. (To measure write latency from a peer or client's perspective, instrument your code to measure the time from when transact is called until the transaction future is realized.)

The [TransactionBytes] Samples metric, again combined with the size of your transactions, characterizes the write throughput of the system. Note that to achieve maximum throughput you will need to pipeline transactions, i.e. arrange to have multiple transactions in flight at the same time.

Custom Monitoring

You can integrate custom monitoring in both transactors and peers by registering a callback. The callback must be on the JVM process's classpath, so you will need to modify your deployment to e.g. include any necessary JARs and resources in the transactor's lib directory of the standard transactor.

Callbacks are registered via the datomic.metricsCallback Java system property in the peer or transactor, or via the metrics-callback setting in the transactor properties file.

Register callback on the transactor

Register the callback for the transactor in the properties file:

metrics-callback=my.ns/handler

Register callback on the peer

Register the callback for the peer with the command line argument:

-Ddatomic.metricsCallback=my.ns/handler

Java callbacks

Java callbacks must be static public methods, taking a single Object argument, named as my.pkg.CallbackClass.handler, e.g.

metrics-callback=my.pkg.CallbackClass.handler

Clojure callbacks

Clojure callbacks must be functions that can take a single argument, named as my.ns/handler, e.g.

metrics-callback=my.ns/handler

Callback argument format

The callback argument format is alpha, and is subject to change. Currently, the argument is a map with keyword keys, and either number or map values. If the value is itself a map, it will contain :lo, :hi, :sum, and :count components, all with numeric values.

The example below shows :HeatbeatMsec with a map value, and :AvailableMB with a numeric value.

{:HeartbeatMsec {:lo 5000, :hi 5001, :sum 55009, :count 11}, :AvailableMB 921.0}

The metrics are the same as those reported to CloudWatch. Note that the set of metrics is dynamic, open, and subject to change over time. A correct callback implementation should be tolerant of

  • metric names it has never seen before
  • metric values in either of the shapes documented above, regardless of what shapes it has seen previously

Configuring CloudWatch Monitoring

The Datomic Pro transactor can be configured to write metrics to CloudWatch. You can find instructions for setting up CloudWatch here.

Use CloudWatch from Anywhere!

Note that you can use CloudWatch with any Datomic Pro transactor, regardless of which storage you are using, and regardless of where your transactor is located. Instructions for configuring CloudWatch for storages other than DynamoDB can be found here.

Available CloudWatch Metrics

Transactor metrics

The Datomic Pro transactor can create at least the following CloudWatch metrics, and may create others as well.

Metric Units Statistic Description
Alarm count sum a problem has occurred (see alarms table below)
AvailableMB MB minimum unused RAM on transactor
ClusterCreateFS msec maximum time to create a "file system" in the storage
CreateEntireIndexMsec msec maximum time to create index, reported at end of indexing job
CreateFulltextIndexMsec msec maximum time to create fulltext portion of index
Datoms count maximum number of unique datoms in the index
DBAddFulltextMsec msec sum time per transaction, to add fulltext
FulltextSegments count sum total number of fulltext segments in the index, per index job
GarbageSegments count sum total number of garbage segments created
GcPauseMsec msec min, avg, max duration of garbage collector induced pauses
HeartbeatMsec msec avg time spacing of heartbeats written by active transactor
HeartbeatMsec msec samples number of heartbeats written by active transactor
HeartbeatMsec msec maximum longest spacing between heartbeats written by active transactor
HeartMonitorMsec msec min, avg, max time spacing of heartbeats observed by standby transactor
IndexDatoms count maximum number of datoms stored by the index, all sorts
IndexSegments count maximum total number of segments in the index
IndexWrites count sum number of segments written by indexing job, reported at end
IndexWriteMsec msec sum time per index segment write
LogIngestBytes bytes maximum in-memory size of log when a database size
LogIngestMsec msec maximum time to play log back when a database start
LogWriteMsec msec sum the time to write to log per transaction batch
Memcache count avg memcache hit ratio, from 0 (no hits) to 1 (all hits)
Memcache count sum total number of cache hits for memcache
Memcache count samples total number of requests to memcache
MemoryIndexMB MB maximum MB of RAM consumed by memory index
MetricReport count sum Count of successful metrics report writes over a 1 min period
ObjectCache count avg object cache hit ratio, from 0 (no hits) to 1 (all hits)
ObjectCache count sum total number of requests to object cache
MemcachedPutMsec msec min, avg, max distribution of successful memcache put latencies
MemcachedPutFailedMsec msec min, avg, max distribution of failed memcache put latencies
RemotePeers count maximum number of remote peers connected
StorageBackoff msec maximum worst case time spent in backoff/retry around calls to storage
StorageBackoff msec samples number of times a storage operation was retried
Storage{Get,Put}Bytes bytes sum throughput of storage operations
Storage{Get,Put}Bytes bytes samples count of storage operations
Storage{Get,Put}Msec msec min, avg, max distribution of storage latencies
TransactionBatch count sum transactions per batch written to storage
TransactionBytes bytes sum volume of transaction data to log, peers
TransactionBytes bytes samples number of transactions
TransactionBytes bytes min, avg, max distribution of transaction sizes
TransactionDatoms count sum datoms per transaction
TransactionApplyMsec msec min, avg, max time for the transactor to apply the transaction data to its in-memory representation of the db
TransactionMsec msec min, avg, max transaction time, includes the time to write to durable storage
Valcache count avg Valcache hit ratio, from 0 (no hits) to 1 (all hits)
Valcache count sum total number of cache hits for Valcache
Valcache count samples total number of requests to Valcache
ValcachePutMsec msec min, avg, max distribution of successful valcache put latencies
ValcachePutFailedMsec msec min, avg, max distribution of failed valcache put latencies
WriterMemcached      

Metrics outside of AWS

If you're analyzing metrics through your own custom callback, or manually calculating statistics from the log, you may need to take additional steps to interpret metrics. The most common case is producing your own average by dividing a sum value for a metric by its count value.

Example using memcache

If you're looking at memcache metrics outside of AWS, you would interpret it as follows:

  • count is the number of requests to memcached.
  • sum is the total number of hits.
  • You would manually calculate misses as count minus sum.
  • You would manually calculate your hit rate as sum divided by count, producing, e.g., a value like 0.95.

Alarms

Cloudwatch allows you to create alarms when a Cloudwatch metric's value it outside of a predetermined range for a defined time period. Alarms can then send a notification to a pager or email, or trigger an Auto Scaling action to expand or contract some resource.

Datomic transactors have a metric named [Alarm]. This metric indicates that something has gone wrong on a transactor, and that human intervention is likely to be necessary. When deploying Datomic in production, you should create a Cloudwatch alarm that notifies you if the Datomic [Alarm] count is ever nonzero.

Each Datomic [Alarm] is accompanied by a more specific metric whose name is prefixed with [Alarm]. You can use these more specific alarms to plan your response to the problem, as shown in the table below.

Alarm Recommended Action
AlarmIndexingJobFailed check for storage problems
AlarmBackPressure let import finish, increase capacity
AlarmUnhandledException contact Datomic support
Alarm{AnythingElse} contact Datomic support

Internal Monitoring

Metrics and events not documented above are internal to Datomic. Internal monitoring is implementation-specific, and subject to change at any time. You should not build systems that rely on internal monitoring. Please ask at the knowledgbase and we will help you find an alternative approach.

The descriptions of internal metrics and events below are provided for informational purposes only.

AlarmStorageSemaphoreTimeout is an internal metric that occurs after a bounding timeout reading or writing storage. AlarmStorageSemaphoreTimeout can occur if Datomic is waiting tens of seconds to hear back from a storage call, and will likely correlate with spikes in documented metrics such as StorageGetMsec or StoragePutMsec. After an AlarmStorageSemaphoreTimeout, check for problems in storage or the network between a Datomic process and storage.