Monitoring overview

The Cloud Pub/Sub API exports metrics via Stackdriver. Stackdriver allows you to create monitoring dashboards and alerts or access the metrics programmatically.

Viewing metrics

Go to Stackdriver in Google Cloud Platform Console to view Stackdriver monitoring dashboards or to define Stackdriver alerts. You can also use the Stackdriver monitoring API to query and view metrics for subscriptions and topics.

Metrics and resource types

• To see the usage metrics that Cloud Pub/Sub reports to Stackdriver, view the Metrics List in the Stackdriver documentation.
• To see the details for the pubsub_topic pubsub_subscription or pubsub_snapshot monitored resource types, view Monitored Resource Types in the Stackdriver documentation.

Monitor topic or subscription quota utilization

You can use the APIs and services quotas dashboard to monitor the current utilization for a given topic or subscription.

Those metrics are:

• pubsub.googleapis.com/topic/byte_cost
• pubsub.googleapis.com/subscription/byte_cost

Note that these metrics are in bytes, whereas quota is measured in kilobytes.

Keeping subscribers healthy

Monitoring the backlog

To ensure that your subscribers are keeping up with the flow of messages, create a dashboard that shows the following metrics, aggregated by resource, for all your subscriptions.:

• subscription/num_undelivered_messages
• subscription/oldest_unacked_message_age

Create alerts that will fire when these values are unusually large in the context of your system. For instance, the absolute number of undelivered messages is not necessarily meaningful. A backlog of a million messages might be acceptable for a million message-per-second subscription, but unacceptable for a one message-per-second subscription.

Symptoms	Problem	Solutions
Both the oldest_unacked_message_age and num_undelivered_messages are growing in tandem.	Subscribers not keeping up with message volume	Add more subscriber threads or processes. Add more subscriber machines or containers. Look for signs of bugs in your code that prevent it from successfully acknowledging messages or processing them in a timely fashion (see Monitoring ack deadline expiration).
If there is a steady, small backlog size combined with a steadily growing oldest_unacked_message_age, there may be a small number of messages that cannot be processed.	Stuck messages	Examine your application logs to understand whether some messages are causing your code to crash. It's unlikely—but possible —that the offending messages are stuck on Cloud Pub/Sub rather than in your client. Raise a support case once you are confident your code successfully processes each message.
The oldest_unacked_message_age exceeds the subscription's message retention duration.	Permanent data loss	Set up an alert that fires well in advance of the subscription's message retention duration lapsing.

Monitoring ack deadline expiration

In order to reduce end-to-end latency of message delivery, Cloud Pub/Sub allows subscriber clients a limited amount of time to acknowledge a given message (known as the "ack deadline") before re-delivering the message. If your subscribers take too long to acknowledge messages, the messages will be re-delivered, resulting in the subscribers seeing duplicate messages. This can happen for a number of reasons:

• Your subscribers are under-provisioned (you need more threads or machines).

• Each message takes longer to process than the message acknowledgement deadline. Google Cloud Platform Client Libraries generally extend the deadline for individual messages up to a configurable maximum. However, a maximum extension deadline is also in effect for the libraries.

• Some messages consistently crash the client.

It can be useful to measure the rate at which subscribers miss the ack deadline. The specific metric depends on the subscription type:

• Pull: subscription/pull_ack_message_operation_count filtered by response_code != "success"

• Streaming Pull: subscription/streaming_pull_ack_message_operation_count filtered by response_code != "success"

• Push: subscription/push_request_count filtered by response_code != "success"

Excessive ack deadline expiration rates can result in costly inefficiencies in your system. You pay for every redelivery and for attempting to process each message repeatedly. Conversely, a small expiration rate (for example, 0.1-1%) might be healthy.

Monitoring push subscriptions

For push subscriptions, you should also monitor these metrics:

• subscription/push_request_count

  Group the metric by response_code and subcription_id. Since Cloud Pub/Sub push subscriptions use response codes as implicit message acknowledgements, it is important to monitor push request response codes. Because push subscriptions exponentially back off when they encounter timeouts or errors, your backlog can grow quickly based on how your endpoint responds.

  Consider setting an alert for high error rates (create a metric filtered by response class), since those rates lead to slower delivery and a growing backlog. However, push request counts are likely to be more useful as a tool for investigating growing backlog size and age.

• subscription/num_outstanding_messages

  Cloud Pub/Sub generally limits the number of outstanding messages. You should aim for fewer than 1000 outstanding messages in most situations. As a rule, the service adjusts the limit based on the overall throughput of the subscription in increments of 1000, once the throughput achieves a rate on the order of ten thousand messages per second. No specific guarantees are made beyond the maximum value, so 1000 is a good guide.

• subscription/push_request_latencies

  This metric helps you understand your push endpoint's response latency distribution. Because of the limit on the number of outstanding messages, endpoint latency affects subscription throughput. If it takes 100 seconds to process each message, your throughput limit is likely to be 10 messages per second.

Keeping publishers healthy

The primary goal of a publisher is to persist message data quickly. Monitor this performance using topic/send_request_count, grouped by response_code. This metric gives you an indication of whether Cloud Pub/Sub is healthy and accepting requests.

A background rate of retryable errors (significantly lower than 1%) should not be a cause for concern, since most GCP Client Libraries retry message failures. You should investigate error rates that are greater than 1%. Because non-retryable codes are handled by your application (rather than the client library), you should examine response codes. If your publisher application does not have a good way of signaling an unhealthy state, consider setting an alert on the send_request_count metric.

It is equally important to track failed publish requests in your publish client. While client libraries generally retry failed requests, they do not guarantee publication. Refer to Publishing messages for ways to detect permanent publish failures when using GCP Client Libraries. At a minimum, your subscriber should log permanent publish errors. If you log those errors to Stackdriver Logging, you can set up a logs-based metric with an alert.