Bufstream

Configure Bufstream for Google Cloud

Enterprise

This product is in beta. For more information about using Bufstream in production, contact us.

This page describes Bufstream's defaults and provides specific recommendations for configuring your Bufstream cluster in Google Cloud Platform (GCP) to get the best combination of price and performance.

Resources and replicas

For initial Bufstream deployments we recommend a deployment size of 3 replicas with a minimum resource request size of 2 cores, 8 GiB of memory. The bufstream Helm chart defaults to these settings, and can be adjusted with the following Helm values:

bufstream:
  deployment:
    # Number of replicas to deploy
    replicaCount: 3
    resources:
      requests:
        cpu: 2
        memory: 8Gi
      limits:
        # Optional
        # cpu: 2
        memory: 8Gi

Network-intensive workloads

Because Bufstream doesn't have a local disk, most I/O occurs over the network to support Kafka produce and fetch requests. Bufstream uses compression, compaction, and caching to minimize the load on the instances. However, because Bufstream needs to write to and read from remote storage, it puts more load on the network than out-of-the-box Kafka.

Horizontal Pod Autoscaler

We recommend configuring your Bufstream deployment with a minimum deployment size of 6 replicas, and a node pool that runs across multiple Availability Zones (AZs). For example, for a node group over 3 AZs, a minimum replica count of 6 keeps cross-AZ network charges down if an instance is unavailable (such as during a deploy). Properly configured clients are directed to Bufstream instances in the same zone. We also recommend maintaining a ratio of 1:4 vCPU to GiB of memory. For example, for a 6 replica deployment, a 1GiB/s workload may demand 16 cores and 64 GiB of memory.

We also recommend autoscaling based on CPU usage with a 50% average usage target. Adjusting the autoscaling threshold impacts the overall cost of your cluster and its ability to respond to bursty workloads effectively. You can configure autoscaling for the Bufstream deployment using the following Helm values:

bufstream:
  deployment:
    autoscaling:
      enabled: true
      # Optional, replicas and target % cpu usage
      minReplicas: 6
      maxReplicas: 18
      targetCPU: "50"

Object storage

Because Bufstream doesn't store data on a local disk, all data from the cluster is written to object storage. Though Bufstream's only requirement is an isolated bucket to write to, we recommend configuring additional settings.

Object retention

Bufstream manages the object lifecycle directly, including deleting expired or compacted objects, so we don't recommend setting a retention policy for the bucket. If you do set a retention policy, it must be longer than the maximum retention of any topic in your Bufstream cluster to guard against data loss.

Bucket lifecycle

To support multi-part uploads, you must configure a lifecycle policy to clean up failed or partially successful uploads. Configuring this policy will stop failed uploads from polluting the bucket and increasing storage costs. We recommend a maximum of 7 days or the topic retention value.

Bucket permissions

For Bufstream to interact with your bucket, you need to update the configuration with the appropriate permissions. Bufstream needs to perform the following bucket operations:

  • storage.objects.create: Create new storage objects
  • storage.objects.get: Retrieve existing storage objects
  • storage.objects.delete: Remove old storage objects according to retention and compaction rules
  • storage.objects.list: View all storage objects to enforce retention and compaction rules
  • storage.multipartUploads.\*: Allow multi-part uploads

Reducing produce latency

Produce latency can be decreased by reducing Bufstream's configured flush interval. Lowering the flush interval will incur greater cost in the form of more frequent writes to object storage. The default value is 100.

configOverrides:
  writer:
    max_flush_interval_ms: 90

Metadata storage (etcd)

Bufstream requires an etcd cluster in which to persist cluster metadata. We recommend configuring etcd with the following settings:

auto-compaction-mode: periodic
auto-compaction-retention: 30s

As etcd is sensitive to disk performance, we recommend using SSD-backed disks.