Fault Tolerance

The Logstash pipeline coordinates the execution of inputs, filters, and outputs. The following schematic sketches the data flow of a pipeline:

input threads | pipeline worker threads

Pipelines in the current release of Logstash process filtering and output in the same thread. Prior to the 2.2 release of Logstash, filtering and output were distinct stages handled by distinct threads. This change to the Logstash architecture improves performance and enables future persistence capabilities. The new pipeline substantially improves thread liveness, decreases resource usage, and increases throughput. The current Logstash pipeline is a micro-batching pipeline, which is inherently more efficient than a one-at-a-time approach. These efficiencies come in many places, two of the more prominent ones being a reduction in contention and a consequent improvement in thread liveness. These efficiencies are especially noticeable on many-core machines.

Each input {} statement in the Logstash configuration file runs in its own thread. Inputs write events to a common Java SynchronousQueue. This queue holds no events, instead transferring each pushed event to a free worker, blocking if all workers are busy. Each pipeline worker thread takes a batch of events off this queue, creating a buffer per worker, runs the batch of events through the configured filters, then runs the filtered events through any outputs. The size of the batch and number of pipeline worker threads are configurable. The following pseudocode illustrates the process flow:

synchronous_queue = SynchronousQueue.new
inputs.each do |input|
  Thread.new do
    input.run(synchronous_queue)
  end
end
num_pipeline_workers.times do
  Thread.new do
    while true
      batch = take_batch(synchronous_queue, batch_size, batch_delay)
      output_batch(filter_batch(batch))
    end
  end
end
wait_for_threads_to_terminate()

There are three configurable options in the pipeline, --pipeline.workers, --pipeline.batch.size, and --pipeline.batch.delay. The --pipeline.workers or -w parameter determines how many threads to run for filter and output processing. If you find that events are backing up, or that the CPU is not saturated, consider increasing the value of this parameter to make better use of available processing power. Good results can even be found increasing this number past the number of available processors as these threads may spend significant time in an I/O wait state when writing to external systems. Legal values for this parameter are positive integers.

The --pipeline.batch.size or -b parameter defines the maximum number of events an individual worker thread collects before attempting to execute filters and outputs. Larger batch sizes are generally more efficient, but increase memory overhead. Some hardware configurations require you to increase JVM heap size by setting the LS_HEAP_SIZE variable to avoid performance degradation with this option. Values of this parameter in excess of the optimum range cause performance degradation due to frequent garbage collection or JVM crashes related to out-of-memory exceptions. Output plugins can process each batch as a logical unit. The Elasticsearch output, for example, issues bulk requests for each batch received. Tuning the -b parameter adjusts the size of bulk requests sent to Elasticsearch.

The --pipeline.batch.delay option rarely needs to be tuned. This option adjusts the latency of the Logstash pipeline. Pipeline batch delay is the maximum amount of time in milliseconds that Logstash waits for new messages after receiving an event in the current pipeline worker thread. After this time elapses, Logstash begins to execute filters and outputs.The maximum time that Logstash waits between receiving an event and processing that event in a filter is the product of the pipeline_batch_delay and pipeline_batch_size settings.

The total number of inflight events is determined by the product of the pipeline_workers and pipeline_batch_size parameters. This product is referred to as the inflight count. Keep the value of the inflight count in mind as you adjust the pipeline_workers and pipeline_batch_size parameters. Pipelines that intermittently receive large events at irregular intervals require sufficient memory to handle these spikes. Configure the LS_HEAP_SIZE option accordingly.

The Logstash defaults are chosen to provide fast, safe performance for most users. To increase performance, increase the number of pipeline workers or the batch size, taking into account the following suggestions:

Measure each change to make sure it increases, rather than decreases, performance. Ensure that you leave enough memory available to cope with a sudden increase in event size. For example, an application that generates exceptions that are represented as large blobs of text. The number of workers may be set higher than the number of CPU cores since outputs often spend idle time in I/O wait conditions.

Threads in Java have names and you can use the jstack, top, and the VisualVM graphical tools to figure out which resources a given thread uses.

On Linux platforms, Logstash labels all the threads it can with something descriptive. For example, inputs show up as [base]<inputname, filter/output workers show up as [base]>workerN, where N is an integer. Where possible, other threads are also labeled to help you identify their purpose.

When tuning Logstash you may have to adjust the heap size. You can use the VisualVM tool to profile the heap. The Monitor pane in particular is useful for checking whether your heap allocation is sufficient for the current workload. The screenshots below show sample Monitor panes. The first pane examines a Logstash instance configured with too many inflight events. The second pane examines a Logstash instance configured with an appropriate amount of inflight events. Note that the specific batch sizes used here are most likely not applicable to your specific workload, as the memory demands of Logstash vary in large part based on the type of messages you are sending.

In the first example we see that the CPU isn’t being used very efficiently. In fact, the JVM is often times having to stop the VM for “full GCs”. Full garbage collections are a common symptom of excessive memory pressure. This is visible in the spiky pattern on the CPU chart. In the more efficiently configured example, the GC graph pattern is more smooth, and the CPU is used in a more uniform manner. You can also see that there is ample headroom between the allocated heap size, and the maximum allowed, giving the JVM GC a lot of room to work with.

Examining the in-depth GC statistics with a tool similar to the excellent VisualGC plugin shows that the over-allocated VM spends very little time in the efficient Eden GC, compared to the time spent in the more resource-intensive Old Gen “Full” GCs.