metron-indexing

Indexing

Introduction

The indexing topology is a topology dedicated to taking the data from the enrichment topology that have been enriched and storing the data in one or more supported indices

• HDFS as rolled text files, one JSON blob per line
• Elasticsearch
• Solr

By default, this topology writes out to both HDFS and one of Elasticsearch and Solr.

Indices are written in batch and the batch size and batch timeout are specified in the Sensor Indexing Configuration via the batchSize and batchTimeout parameters. These configs are variable by sensor type.

Indexing Architecture

The indexing topology is extremely simple. Data is ingested into kafka and sent to

• An indexing bolt configured to write to either elasticsearch or Solr
• An indexing bolt configured to write to HDFS under /apps/metron/enrichment/indexed

By default, errors during indexing are sent back into the indexing kafka queue so that they can be indexed and archived.

Sensor Indexing Configuration

The sensor specific configuration is intended to configure the indexing used for a given sensor type (e.g. snort).

Just like the global config, the format is a JSON stored in zookeeper and on disk at $METRON_HOME/config/zookeeper/indexing. Within the sensor-specific configuration, you can configure the individual writers. The writers currently supported are:

• elasticsearch
• hdfs
• solr

Depending on how you start the indexing topology, it will have either elasticsearch or solr and hdfs writers running.

The configuration for an individual writer-specific configuration is a JSON map with the following fields:

• index : The name of the index to write to (defaulted to the name of the sensor).
• batchSize : The size of the batch that is written to the indices at once. Defaults to 1 (no batching).
• batchTimeout : The timeout after which a batch will be flushed even if batchSize has not been met. Optional. If unspecified, or set to 0, it defaults to a system-determined duration which is a fraction of the Storm parameter topology.message.timeout.secs. Ignored if batchSize is 1, since this disables batching.
• enabled : Whether the writer is enabled (default true).

Meta Alerts

Alerts can be grouped, after appropriate searching, into a set of alerts called a meta alert. A meta alert is useful for maintaining the context of searching and grouping during further investigations. Standard searches can return meta alerts, but grouping and other aggregation or sorting requests will not, because there's not a clear way to aggregate in many cases if there are multiple alerts contained in the meta alert. All meta alerts will have the source type of metaalert, regardless of the contained alert's origins.

Elasticsearch

Metron comes with built-in templates for the default sensors for Elasticsearch. When adding a new sensor, it will be necessary to add a new template defining the output fields appropriately. In addition, there is a requirement for a field alert of type nested for Elasticsearch 2.x installs. This is detailed at Using Metron with Elasticsearch 2.x

Indexing Configuration Examples

For a given sensor, the following scenarios would be indicated by the following cases:

Base Case

{
}

or no file at all.

• elasticsearch writer
  • enabled
  • batch size of 1
  • batch timeout system default
  • index name the same as the sensor
• hdfs writer
  • enabled
  • batch size of 1
  • batch timeout system default
  • index name the same as the sensor

If a writer config is unspecified, then a warning is indicated in the Storm console. e.g.: WARNING: Default and (likely) unoptimized writer config used for hdfs writer and sensor squid

Fully specified

{
   "elasticsearch": {
      "index": "foo",
      "batchSize" : 100,
      "batchTimeout" : 0,
      "enabled" : true 
    },
   "hdfs": {
      "index": "foo",
      "batchSize": 1,
      "batchTimeout" : 0,
      "enabled" : true
    }
}

• elasticsearch writer
  • enabled
  • batch size of 100
  • batch timeout system default
  • index name of "foo"
• hdfs writer
  • enabled
  • batch size of 1
  • batch timeout system default
  • index name of "foo"

HDFS Writer turned off

{
   "elasticsearch": {
      "index": "foo",
      "enabled" : true 
    },
   "hdfs": {
      "index": "foo",
      "batchSize": 100,
      "batchTimeout" : 0,
      "enabled" : false
    }
}

• elasticsearch writer
  • enabled
  • batch size of 1
  • batch timeout system default
  • index name of "foo"
• hdfs writer
  • disabled

Updates to Indexed Data

There are clear usecases where we would want to incorporate the capability to update indexed data. Thus far, we have limited capabilities provided to support this use-case:

• Updates to the random access index (e.g. Elasticsearch and Solr) should be supported
• Updates to the cold storage index (e.g. HDFS) is not supported currently, however to support the batch use-case updated documents will be provided in a NoSQL write-ahead log (e.g. a HBase table) and an Java API will be provided to retrieve those updates scalably (i.e. a scan-free architecture).

Put simply, the random access index will be always up-to-date, but the HDFS index will need to be joined to the NoSQL write-ahead log to get current updates.

The IndexDao Abstraction

The indices mentioned above as part of Update should be pluggable by the developer so that new write-ahead logs or real-time indices can be supported by providing an implementation supporting the data access patterns.

To support a new index, one would need to implement the org.apache.metron.indexing.dao.IndexDao abstraction and provide update and search capabilities. IndexDaos may be composed and updates will be performed in parallel. This enables a flexible strategy for specifying your backing store for updates at runtime. For instance, currently the REST API supports the update functionality and may be configured with a list of IndexDao implementations to use to support the updates.

Updates with the IndexDao.update method replace the current object with the new object. For partial updates, use IndexDao.patch instead.

The HBaseDao

Updates will be written to HBase. The key structure includes the GUID and sensor type and for each new version, a new column is created with value as the message.

The HBase table and column family are configured via fields in the global configuration.

update.hbase.table

The HBase table to use for message updates.

update.hbase.cf

The HBase column family to use for message updates.

The MetaAlertDao

The goal of meta alerts is to be able to group together a set of alerts while being able to transparently perform actions like searches, as if meta alerts were normal alerts. org.apache.metron.indexing.dao.MetaAlertDao extends IndexDao and enables several features:

• the ability to get all meta alerts associated with an alert
• creation of a meta alert
• adding alerts to a meta alert
• removing alerts from a meta alert
• changing a meta alert's status

The implementation of this is to denormalize the relationship between alerts and meta alerts, and store alerts as a nested field within a meta alert. The use of nested fields is to avoid the limitations of parent-child relationships (one-to-many) and merely linking by IDs (which causes issues with pagination as a result of being unable to join indices). A list of containing meta alerts is stored on an alert for the purpose of keeping source alerts and alerts contained in meta alerts in sync.

The search functionality of IndexDao is wrapped by the MetaAlertDao in order to provide both regular and meta alerts side-by-side with sorting. The updating capabilities are similarly wrapped, in order to ensure updates are carried through both the alerts and associated meta alerts. Both of these functions are handled under the hood.

In addition, API endpoints have been added to expose the features listed above. The denormalization handles the case of going from meta alert to alert automatically.

Notes on Performance Tuning

Default installed Metron is untuned for production deployment. By far and wide, the most likely piece to require TLC from a performance perspective is the indexing layer. An index that does not keep up will back up and you will see errors in the kafka bolt. There are a few knobs to tune to get the most out of your system.

Kafka Queue

The indexing kafka queue is a collection point from the enrichment topology. As such, make sure that the number of partitions in the kafka topic is sufficient to handle the throughput that you expect.

Indexing Topology

The indexing topology as started by the $METRON_HOME/bin/start_elasticsearch_topology.sh or $METRON_HOME/bin/start_solr_topology.sh script uses a default of one executor per bolt. In a real production system, this should be customized by modifying the flux file in $METRON_HOME/flux/indexing/remote.yaml.

• Add a parallelism field to the bolts to give Storm a parallelism hint for the various components. Give bolts which appear to be bottlenecks (e.g. the indexing bolt) a larger hint.
• Add a parallelism field to the kafka spout which matches the number of partitions for the enrichment kafka queue.
• Adjust the number of workers for the topology by adjusting the topology.workers field for the topology.

Finally, if workers and executors are new to you or you don't know where to modify the flux file, the following might be of use to you:

• Understanding the Parallelism of a Storm Topology
• Flux Docs

Zeppelin Notebooks

Zeppelin notebooks can be added to /src/main/config/zeppelin/ (and subdirectories can be created for organization). The placed files must be .json files and be named appropriately. These files must be added to the metron.spec file and the RPMs rebuilt to be available to be loaded into Ambari.

The notebook files will be found on the server in $METRON_HOME/config/zeppelin

The Ambari Management Pack has a custom action to load these templates, ZEPPELIN_DASHBOARD_INSTALL, that will import them into Zeppelin.