Skip to main content

Persistent & streaming log template miner

Project description



Drain3 is an online log template miner that can extract templates (clusters) from a stream of log messages in a timely manner. It employs a parse tree with fixed depth to guide the log group search process, which effectively avoids constructing a very deep and unbalanced tree.

Drain3 continuously learns on-the-fly and extracts log templates from raw log entries.


For the input:

connected to
connected to
Hex number 0xDEADBEAF
user davidoh logged in
user eranr logged in

Drain3 extracts the following templates:

ID=1     : size=2         : connected to <:IP:>
ID=2     : size=1         : Hex number <:HEX:>
ID=3     : size=2         : user <:*:> logged in

Full sample program output:

Starting Drain3 template miner
Checking for saved state
Saved state not found
Drain3 started with 'FILE' persistence
Starting training mode. Reading from std-in ('q' to finish)
> connected to
Saving state of 1 clusters with 1 messages, 528 bytes, reason: cluster_created (1)
{"change_type": "cluster_created", "cluster_id": 1, "cluster_size": 1, "template_mined": "connected to <:IP:>", "cluster_count": 1}
Parameters: [ExtractedParameter(value='', mask_name='IP')]
> connected to
{"change_type": "none", "cluster_id": 1, "cluster_size": 2, "template_mined": "connected to <:IP:>", "cluster_count": 1}
Parameters: [ExtractedParameter(value='', mask_name='IP')]
> Hex number 0xDEADBEAF
Saving state of 2 clusters with 3 messages, 584 bytes, reason: cluster_created (2)
{"change_type": "cluster_created", "cluster_id": 2, "cluster_size": 1, "template_mined": "Hex number <:HEX:>", "cluster_count": 2}
Parameters: [ExtractedParameter(value='0xDEADBEAF', mask_name='HEX')]
> user davidoh logged in
Saving state of 3 clusters with 4 messages, 648 bytes, reason: cluster_created (3)
{"change_type": "cluster_created", "cluster_id": 3, "cluster_size": 1, "template_mined": "user davidoh logged in", "cluster_count": 3}
Parameters: []
> user eranr logged in
Saving state of 3 clusters with 5 messages, 644 bytes, reason: cluster_template_changed (3)
{"change_type": "cluster_template_changed", "cluster_id": 3, "cluster_size": 2, "template_mined": "user <:*:> logged in", "cluster_count": 3}
Parameters: [ExtractedParameter(value='eranr', mask_name='*')]
> q
Training done. Mined clusters:
ID=1     : size=2         : connected to <:IP:>
ID=2     : size=1         : Hex number <:HEX:>
ID=3     : size=2         : user <:*:> logged in

This project is an upgrade of the original Drain project by LogPAI from Python 2.7 to Python 3.6 or later with additional features and bug-fixes.

Read more information about Drain from the following paper:

A Drain3 use case is presented in this blog post: Use open source Drain3 log-template mining project to monitor for network outages .

New features

  • Persistence. Save and load Drain state into an Apache Kafka topic, Redis or a file.
  • Streaming. Support feeding Drain with messages one-be-one.
  • Masking. Replace some message parts (e.g numbers, IPs, emails) with wildcards. This improves the accuracy of template mining.
  • Packaging. As a pip package.
  • Configuration. Support for configuring Drain3 using an .ini file or a configuration object.
  • Memory efficiency. Decrease the memory footprint of internal data structures and introduce cache to control max memory consumed (thanks to @StanislawSwierc)
  • Inference mode. In case you want to separate training and inference phase, Drain3 provides a function for fast matching against already-learned clusters (templates) only, without the usage of regular expressions.
  • Parameter extraction. Accurate extraction of the variable parts from a log message as an ordered list, based on its mined template and the defined masking instructions (thanks to @Impelon).

Expected Input and Output

Although Drain3 can be ingested with full raw log message, template mining accuracy can be improved if you feed it with only the unstructured free-text portion of log messages, by first removing structured parts like timestamp, hostname. severity, etc.

The output is a dictionary with the following fields:

  • change_type - indicates either if a new template was identified, an existing template was changed or message added to an existing cluster.
  • cluster_id - Sequential ID of the cluster that the log belongs to.
  • cluster_size- The size (message count) of the cluster that the log belongs to.
  • cluster_count - Count clusters seen so far.
  • template_mined- the last template of above cluster_id.


Drain3 is configured using configparser. By default, config filename is drain3.ini in working directory. It can also be configured passing a TemplateMinerConfig object to the TemplateMiner constructor.

Primary configuration parameters:

  • [DRAIN]/sim_th - similarity threshold. if percentage of similar tokens for a log message is below this number, a new log cluster will be created (default 0.4)
  • [DRAIN]/depth - max depth levels of log clusters. Minimum is 2. (default 4)
  • [DRAIN]/max_children - max number of children of an internal node (default 100)
  • [DRAIN]/max_clusters - max number of tracked clusters (unlimited by default). When this number is reached, model starts replacing old clusters with a new ones according to the LRU cache eviction policy.
  • [DRAIN]/extra_delimiters - delimiters to apply when splitting log message into words (in addition to whitespace) ( default none). Format is a Python list e.g. ['_', ':'].
  • [MASKING]/masking - parameters masking - in json format (default "")
  • [MASKING]/mask_prefix & [MASKING]/mask_suffix - the wrapping of identified parameters in templates. By default, it is < and > respectively.
  • [SNAPSHOT]/snapshot_interval_minutes - time interval for new snapshots (default 1)
  • [SNAPSHOT]/compress_state - whether to compress the state before saving it. This can be useful when using Kafka persistence.


This feature allows masking of specific variable parts in log message with keywords, prior to passing to Drain. A well-defined masking can improve template mining accuracy.

Template parameters that do not match any custom mask in the preliminary masking phase are replaced with <*> by Drain core.

Use a list of regular expressions in the configuration file with the format {'regex_pattern', 'mask_with'} to set custom masking.

For example, following masking instructions in drain3.ini will mask IP addresses and integers:

masking = [
          {"regex_pattern":"((?<=[^A-Za-z0-9])|^)(\\d{1,3}\\.\\d{1,3}\\.\\d{1,3}\\.\\d{1,3})((?=[^A-Za-z0-9])|$)", "mask_with": "IP"},
          {"regex_pattern":"((?<=[^A-Za-z0-9])|^)([\\-\\+]?\\d+)((?=[^A-Za-z0-9])|$)", "mask_with": "NUM"},


The persistence feature saves and loads a snapshot of Drain3 state in a (compressed) json format. This feature adds restart resiliency to Drain allowing continuation of activity and maintain learned knowledge across restarts.

Drain3 state includes the search tree and all the clusters that were identified up until snapshot time.

The snapshot also persist number of log messages matched each cluster, and it's cluster_id.

An example of a snapshot:

  "clusters": [
      "cluster_id": 1,
      "log_template_tokens": [
      "py/object": "drain3_core.LogCluster",
      "size": 2
      "cluster_id": 2,
      "log_template_tokens": [
      "py/object": "drain3_core.LogCluster",
      "size": 1

This example snapshot persist two clusters with the templates:

["aa", "aa", "<*>"] - occurs twice

["My", "IP", "is", "<IP>"] - occurs once

Snapshots are created in the following events:

  • cluster_created - in any new template
  • cluster_template_changed - in any update of a template
  • periodic - after n minutes from the last snapshot. This is intended to save cluster sizes even if no new template was identified.

Drain3 currently supports the following persistence modes:

  • Kafka - The snapshot is saved in a dedicated topic used only for snapshots - the last message in this topic is the last snapshot that will be loaded after restart. For Kafka persistence, you need to provide: topic_name. You may also provide other kwargs that are supported by kafka.KafkaConsumer and kafka.Producer e.g bootstrap_servers to change Kafka endpoint (default is localhost:9092).

  • Redis - The snapshot is saved to a key in Redis database (contributed by @matabares).

  • File - The snapshot is saved to a file.

  • Memory - The snapshot is saved an in-memory object.

  • None - No persistence.

Drain3 persistence modes can be easily extended to another medium / database by inheriting the PersistenceHandler class.

Training vs. Inference modes

In some use-cases, it is required to separate training and inference phases.

In training phase you should call template_miner.add_log_message(log_line). This will match log line against an existing cluster (if similarity is above threshold) or create a new cluster. It may also change the template of an existing cluster.

In inference mode you should call template_miner.match(log_line). This will match log line against previously learned clusters only. No new clusters are created and templates of existing clusters are not changed. Match to existing cluster has to be perfect, otherwise None is returned. You can use persistence option to load previously trained clusters before inference.

Memory efficiency

This feature limits the max memory used by the model. It is particularly important for large and possibly unbounded log streams. This feature is controlled by the max_clusters​ parameter, which sets the max number of clusters/templates trarcked by the model. When the limit is reached, new templates start to replace the old ones according to the Least Recently Used (LRU) eviction policy. This makes the model adapt quickly to the most recent templates in the log stream.

Parameter Extraction

Drain3 supports retrieving an ordered list of variables in a log message, after its template was mined. Each parameter is accompanied by the name of the mask that was matched, or * for the catch-all mask.

Parameter extraction is performed by generating a regular expression that matches the template and then applying it on the log message. When exact_matching is enabled (by default), the generated regex included the regular expression defined in relevant masking instructions. If there are multiple masking instructions with the same name, either match can satisfy the regex. It is possible to disable exact matching so that every variable is matched against a non-whitespace character sequence. This may improve performance on expanse of accuracy.

Parameter extraction regexes generated per template are cached by default, to improve performance. You can control cache size with the MASKING/parameter_extraction_cache_capacity configuration parameter.

Sample usage:

result = template_miner.add_log_message(log_line)
params = template_miner.extract_parameters(
    result["template_mined"], log_line, exact_matching=True)

For the input "user johndoe logged in 11 minuts ago", the template would be:

"user <:*:> logged in <:NUM:> minuts ago"

... and the extracted parameters:

  ExtractedParameter(value='johndoe', mask_name='*'), 
  ExtractedParameter(value='11', mask_name='NUM')


Drain3 is available from PyPI. To install use pip:

pip3 install drain3

Note: If you decide to use Kafka or Redis persistence, you should install relevant client library explicitly, since it is declared as an extra (optional) dependency, by either:

pip3 install kafka-python

-- or --

pip3 install redis


In order to run the examples directly from the repository, you need to install dependencies. You can do that using * pipenv* by executing the following command (assuming pipenv already installed):

python3 -m pipenv sync

Example 1 - drain_stdin_demo

Run examples/ from the root folder of the repository by:

python3 -m pipenv run python -m examples.drain_stdin_demo

This example uses Drain3 on input from stdin and persist to either Kafka / file / no persistence.

Change persistence_type variable in the example to change persistence mode.

Enter several log lines using the command line. Press q to end online learn-and-match mode.

Next, demo goes to match (inference) only mode, in which no new clusters are trained and input is matched against previously trained clusters only. Press q again to finish execution.

Example 2 - drain_bigfile_demo

Run examples/drain_bigfile_demo from the root folder of the repository by:

python3 -m pipenv run python -m examples.drain_bigfile_demo

This example downloads a real-world log file (of an SSH server) and process all lines, then prints result clusters, prefix tree and performance statistics.

Sample config file

An example drain3.ini file with masking instructions can be found in the examples folder as well.


Our project welcomes external contributions. Please refer to for further details.

Change Log

  • Fixed possible DivideByZero error when the profiler is enabled - Issue #65.
  • Fixed compatibility issue with Python 3.10 caused by removal of KeysView.
  • Added support for accurate log message parameter extraction in a new function - extract_parameters(). The function get_parameter_list() is deprecated (Thanks to @Impelon).
  • Refactored AbstractMaskingInstruction as a base class for RegexMaskingInstruction, allowing to introduce other types of masking mechanisms.
  • Added an option full_search_strategy option in TemplateMiner.match() and Drain.match(). See more info at Issue #48.
  • Added an option to disable parameterization of tokens that contains digits in configuration: TemplateMinerConfig.parametrize_numeric_tokens
  • Loading Drain snapshot now only restores clusters state and not configuration parameters. This improves backwards compatibility when introducing new Drain configuration parameters.
  • Fixed bug in original Drain: log clusters were created multiple times for log messages with fewer tokens than max_node_depth.
  • Changed depth property name to a more descriptive name max_node_depth as Drain always subtracts 2 of depth argument value. Also added log_cluster_depth property to reflect original value of depth argument (Breaking Change).
  • Restricted depth param to minimum sensible value of 3.
  • Added log cluster count to nodes in Drain.print_tree()
  • Added optional log cluster details to Drain.print_tree()
  • Added: TemplateMiner.match() function for fast matching against existing clusters only.
  • Added: TemplateMiner.get_parameter_list() function to extract template parameters for raw log message (thanks to * @cwyalpha*)
  • Added option to customize mask wrapper - Instead of the default <*>, <NUM> etc, you can select any wrapper prefix or suffix by overriding TemplateMinerConfig.mask_prefix and TemplateMinerConfig.mask_prefix
  • Fixed: config .ini file is always read from same folder as source file in demos in tests (thanks @RobinMaas95)
  • Fixed: comparison of type int with type str in function add_seq_to_prefix_tree #28 (bug introduced at v0.9.1)
  • Updated jsonpickle version
  • Keys id_to_cluster dict are now persisted by jsonpickle as int instead of str to avoid keys type conversion on load snapshot which caused some issues.
  • Added cachetools dependency to
  • Added option to configure TemplateMiner using a configuration object (without .ini file).
  • Support for print_tree() to a file/stream.
  • Added MemoryBufferPersistence
  • Added unit tests for state save/load.
  • Bug fix: missing type-conversion in state loading, introduced in v0.9.0
  • Refactor: Drain prefix tree keys are now of type str also for 1st level (was int before), for type consistency.
  • Decrease memory footprint of the main data structures.
  • Added max_clusters option to limit the number of tracked clusters.
  • Changed cluster identifier type from str to int
  • Added more unit tests and CI
  • Added extra_delimiters configuration option to Drain
  • Profiler improvements
  • Masking speed improvement
  • Fix: profiler state after load from snapshot
  • Fixed snapshot backward compatibility to v0.7.9
  • Bugfix in profiling configuration read
  • Added time profiling support (disabled by default)
  • Added cluster ID to snapshot reason log (credit: @boernd)
  • Minor Readability and documentation improvements in Drain
  • Fix: KafkaPersistence now accepts also bootstrap_servers as kwargs.
  • Using kafka-python package instead of kafka (newer).
  • Added support for specifying additional configuration as kwargs in Kafka persistence handler.
  • Corrected default Drain config values.
  • Improvement in config file handling (Note: new sections were added instead of DEFAULT section)
  • Made Kafka and Redis optional requirements

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

drain3-0.9.11.tar.gz (28.0 kB view hashes)

Uploaded source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page