Tune Horizontal Autoscaling for streaming pipelines

In streaming pipelines with a high volume of input data, there is generally a tradeoff between cost and latency. To maintain low latency, Dataflow must add workers as the volume of traffic increases. Another factor is how quickly the pipeline should scale up or down in response to changes in the input data rate.

The Dataflow autoscaler has default settings that are suitable for many workloads. However, you might want to tune this behavior for your particular scenario. For example, a higher average latency might be acceptable in order to reduce costs, or you might want Dataflow to scale up more quickly in response to spikes in traffic.

To optimize Horizontal Autoscaling, you can adjust the following parameters:

• Autoscaling range: The minimum and maximum number of workers to allocate.
• Worker utilization hint: The target CPU utilization for workers.

Set the autoscaling range

When you create a new streaming job, you can set the initial number of workers and the maximum number of workers. To do so, specify the following pipeline options:

For streaming jobs that use Streaming Engine, the --maxNumWorkers flag is optional. The default is 100. For streaming jobs not using Streaming Engine, --maxNumWorkers is required when Horizontal Autoscaling is enabled.

The starting value of --maxNumWorkers also determines how many Persistent Disks are allocated for the job. Pipelines are deployed with a fixed pool of Persistent Disks, equal in number to --maxNumWorkers. During streaming, Persistent Disks are redistributed such that each worker gets an equal number of attached disks.

If you set --maxNumWorkers, ensure the value provides enough disks for your pipeline. Consider future growth when setting the initial value. For information about Persistent Disk performance, see Configure your Persistent Disk and VMs. Dataflow bills for Persistent Disk usage and has Compute Engine quotas, including Persistent Disk quotas.

By default, the minimum number of workers is 1 for streaming jobs that use Streaming Engine, and (maxNumWorkers/15), rounded up, for jobs that don't use Streaming Engine.

Update the autoscaling range

For jobs that use Streaming Engine, you can adjust the minimum and maximum number workers, without stopping or replacing the job. To adjust these values, use an in-flight job update. Update the following job options:

• --min-num-workers: the minimum number of workers.

• --max-num-workers: the maximum number of workers.

gcloud dataflow jobs update-options \
  --region=REGION \
  --min-num-workers=MINIMUM_WORKERS \
  --max-num-workers=MAXIMUM_WORKERS \
  JOB_ID

PUT https://dataflow.googleapis.com/v1b3/projects/PROJECT_ID/locations/REGION/jobs/JOB_ID?updateMask=runtime_updatable_params.max_num_workers,runtime_updatable_params.min_num_workers
{
  "runtime_updatable_params": {
    "min_num_workers": MINIMUM_WORKERS,
    "max_num_workers": MAXIMUM_WORKERS
  }
}

PUT https://dataflow.googleapis.com/v1b3/projects/my_project/locations/us-central1/jobs/job1?updateMask=runtime_updatable_params.min_num_workers
{
  "runtime_updatable_params": {
    "min_num_workers": 5
  }
}

For jobs that don't use Streaming Engine, you can replace the existing job with an updated value of maxNumWorkers.

If you update a streaming job that is not using Streaming Engine, the updated job has Horizontal Autoscaling disabled by default. To keep autoscaling enabled, specify --autoscalingAlgorithm and --maxNumWorkers for the updated job.

Set the worker utilization hint

Dataflow uses the average CPU utilization as a signal for when to apply Horizontal Autoscaling. By default, Dataflow sets a target CPU utilization of 0.8. When utilization falls outside of this range, Dataflow might add or remove workers.

For more control over the autoscaling behavior, you can set the target CPU utilization to a value in the range [0.1, 0.9].

• Set a lower CPU utilization value if you want to achieve lower peak latencies. A lower value allows Dataflow to scale out more aggressively in response to growing worker utilization, and to downscale more conservatively to improve stability. A lower value also provides more headroom when the pipeline is running at a steady state, generally resulting in a lower tail latency. (Tail latency measures the longest wait times before a new record is processed.)

• Set a higher value if you want to save resources and keep costs lower when traffic spikes. A higher value prevents excessive upscaling, at the expense of higher latency.

To configure the utilization hint when you run a non-template job, set the worker_utilization_hint service option. For a template job, update the utilization hint instead, as service options are not supported.

The following example shows how to use worker_utilization_hint:

--dataflowServiceOptions=worker_utilization_hint=TARGET_UTILIZATION

--dataflow_service_options=worker_utilization_hint=TARGET_UTILIZATION

--dataflow_service_options=worker_utilization_hint=TARGET_UTILIZATION

For new pipelines, we recommend that you test under realistic loads using the default setting. Then evaluate the autoscaling behavior as it applies to your pipeline and make adjustments as needed.

The utilization hint is just one factor that Dataflow uses when deciding whether to scale workers. Other factors such as backlog and available keys can override the hint value. Also, the hint is not a strict target. The autoscaler tries to keep CPU utilization within the range of the hint value, but the aggregated utilization metric might be higher or lower. For more information, see Streaming autoscaling heuristics.

Update the utilization hint

To update the utilization hint while a job is running, perform an in-flight update as follows:

gcloud dataflow jobs update-options \
  --region=REGION \
  --worker-utilization-hint=TARGET_UTILIZATION \
  JOB_ID

gcloud dataflow jobs update-options \
  --unset-worker-utilization-hint \
  --region=REGION \
  --project=PROJECT_ID \
  JOB_ID

PUT https://dataflow.googleapis.com/v1b3/projects/PROJECT_ID/locations/REGION/jobs/JOB_ID?updateMask=runtime_updatable_params.worker_utilization_hint
{
  "runtime_updatable_params": {
    "worker_utilization_hint": TARGET_UTILIZATION
  }
}

Streaming autoscaling heuristics

For streaming pipelines, the objective of Horizontal Autoscaling is to minimize backlog while maximizing worker utilization and throughput and to react quickly to spikes in load.

Dataflow takes several factors into account when autoscaling, including:

• Backlog. The estimated backlog time is calculated from the throughput and the backlog bytes still to be processed from the input source. A pipeline is considered backlogged when the estimated backlog time stays above 15 seconds.

• Target CPU utilization. The default target for average CPU utilization is 0.8. You can override this value.

• Available keys. Keys are the fundamental unit of parallelism in Dataflow.

In some cases, Dataflow uses the following factors in autoscaling decisions. If these factors are used for your job, you can see that information in the Autoscaling metrics tab.

• Key-based throttling uses the number of processing keys received by the job to calculate the cap for user workers, because each key can only be processed by one worker at a time.

• Downscale dampening. If Dataflow detects that unstable autoscaling decisions have occurred, it slows the rate of downscaling in order to improve stability.

• CPU-based upscale uses high CPU utilization as an upscaling criteria.

• For streaming jobs that don't use Streaming Engine, scaling might be constrained by the number of Persistent Disks. For more information, see Set the autoscaling range.

Upscaling. If a streaming pipeline remains backlogged with sufficient parallelism on the workers for several minutes, Dataflow scales up. Dataflow attempts to clear the backlog within approximately 150 seconds of scaling up, given the current throughput per worker. If there is backlog but the worker doesn't have enough parallelism for additional workers, the pipeline does not scale up. (Scaling the number of workers beyond the number of keys available for parallel processing doesn't help process the backlog faster.)

Downscaling When the autoscaler makes a downscaling decision, backlog is the highest priority factor. The autoscaler aims for a backlog of not more than 15 seconds. If the backlog drops below 10 seconds, and average worker utilization is below the CPU utilization target, then Dataflow scales down. As long as the backlog is acceptable, the autoscaler tries to maintain CPU utilization close to the target CPU utilization. If utilization is already sufficiently close to the target, however, the autoscaler might keep the number of workers unchanged, because each downscaling step has a cost.

Streaming Engine also uses a predictive autoscaling technique based on timer backlog. Unbounded data in a streaming pipeline is divided into windows grouped by timestamps. At the end of a window, timers fire for each key being processed in that window. The firing of a timer indicates that the window is expired for a given key. Streaming Engine can measure the timer backlog and predict how many timers will fire at the end of a window. By using the timer backlog as a signal, Dataflow can estimate the amount of processing that must happen when future timers fire. Based on the estimated future load, Dataflow autoscales in advance to meet expected demand.

Metrics

To find the current autoscale limits for a job, query the following metrics:

• job/max_worker_instances_limit: Maximum number of workers.
• job/min_worker_instances_limit: Minimum number of workers.

To get information about worker utilization, query the following metrics:

• job/aggregated_worker_utilization: The aggregated worker utilization.
• job/worker_utilization_hint: The current worker utilization hint.

To get insights into the behavior of the autoscaler, query the following metric:

• job.worker_utilization_hint_is_actively_used: Indicates whether the autoscaler is actively using the worker utilization hint. If other factors override the hint when this metric is sampled, the value is false.
• job/horizontal_worker_scaling: Describes the decisions taken by the autoscaler. This metric contains the following labels:
  • direction: Specifies whether the autoscaler scaled up, scaled down, or took no action.
  • rationale: Specifies the rationale for the autoscaler's decision.

For more information, see Cloud Monitoring metrics. These metrics are also displayed in the autoscaling monitoring charts.

What's next

• Monitor Dataflow autoscaling
• Troubleshoot Dataflow autoscaling