Service Performance

Unreliable or slow services can lead to a poor user experience for your customers. With Pixie, you can get immediate visibility into the health of your services, without the need for manual instrumentation.

Pixie automatically captures all network traffic in your cluster using eBPF, a low-level Linux tracing technology. Messages of a supported protocol type, such as HTTP2/gRPC, are parsed and paired with their responses, making latency, error, and throughput information immediately available after installing Pixie.

This tutorial will demonstrate how to use Pixie to see:

• The flow of HTTP traffic between the services in your cluster.
• Latency per service.
• Latency per service endpoint.
• A sample of the slowest requests for an individual service.

If you're interested in troubleshooting HTTP errors, check out the Request Tracing tutorial.

1. You will need a Kubernetes cluster with Pixie installed. If you do not have a cluster, you can create a minikube cluster and install Pixie using one of our install guides.

2. You will need to install the demo microservices application, using Pixie's CLI:

• Install the Pixie CLI.
• Run px demo deploy px-sock-shop to install Weavework's Sock Shop demo app.
• Run kubectl get pods -n px-sock-shop to make sure all pods are ready before proceeding. The demo app can take up to 5 minutes to deploy.

When debugging issues with microservices, it helps to start at a high-level view, like a service map, and then drill down into the problem service(s).

For a global view of the services in your cluster, we'll use the px/cluster script:

1. Open the Live UI and select px/cluster from the script drop-down menu at the top.

This script shows a graph of the HTTP traffic between the services in your cluster, along with latency, error, and throughput rate per service.

Click the Kubernetes icon for a shortcut to the `px/cluster` script.

Hover over an edge for latency, error and throughput. Thicker lines indicate more traffic.

Drag the 3-dot column divider to expand the column.

2. Scroll down to the Services table.

This table contains latency, error and throughput rate for all HTTP traffic. The INBOUND_THROUGHPUT and OUTBOUND_THROUGHPUT columns reflect all traced network traffic (not just HTTP) for the service.

Let's figure out which service is the slowest.

3. Click the LATENCY column title to sort the services by latency.

It’s good to check multiple percentiles for latency, not just the average, in order to get a better picture of the overall distribution.

4. Expand the LATENCY column by dragging the 3-dot column header divider.

This script represents service latency with a box & whisker plot.

5. Click the vertical quantile lines on the box plot to switch the latency display between the P50, P90 and P99 quantile values.

The LATENCY column will resort itself and the column title will update to reflect the selected quantile.

A high P50 latency value for the front-end service indicates that this is general performance degradation, rather than an issue with a specific request.

Once we have identified a service we are interested in investigating further, we will want to drill down into its detailed latency information.

Pixie's UI makes it easy to quickly navigate between Kubernetes resources. Clicking on any pod, node, service, or namespace name in the UI will open a script showing a high-level overview for that entity.

6. From the SERVICE column in the Services table, click on the px-sock-shop/front-end service.

This will open the px/service script with the service argument pre-filled with the name of the service you selected.

The px/service script shows the latency, error, and throughput over time for all HTTP requests for the service.

Click a legend item to highlight those specific results. Click the item a second time to show all results.

Drag your mouse across the graph to see the values at particular timestamps.

Modify the start_time to change the time window for the results (e.g -30m, -1h).

This view shows us that the service's latency values have been consistent over the selected time window.

7. Scroll down to the Sample of Slow Requests table and expand the REQ_PATH column.

If this service handles multiple kinds of requests, this table can help identify if there is a particular request type that is much slower.

This table shows individual requests, so we will see the full path with URL parameters filled in (for example, /restaurants/123). However, Pixie makes it possible to drill down into individual logical endpoints (for example, /restaurants/*).

Request latency can vary greatly by endpoint, especially if one of the requests is more database intensive. However, when there are wildcards (URL parameters) in your request paths, it can be difficult to drill down into a particular endpoint.

Pixie can cluster HTTP requests by logical endpoint, substituting a * for the parameters in your requests. For example, the following two requests:

/restaurants/0123550/reviews/239487345/author /restaurants/3485799/reviews/394853457/author

would be clustered together into the logical endpoint:

/restaurants/*/reviews/*/author

Let's look at latency by logical service endpoint:

1. Select pxbeta/service_endpoints from the script drop-down menu (note: this is a Beta script).

2. Select the drop-down arrow next to the service argument, type px-sock-shop/catalogue, and press Enter to re-run the script.

This script shows latency, error and throughput per logical endpoint for the given service.

3. Click on catalog/* in the Endpoints table to see an overview of that individual endpoint with a sample of slow requests.

This tutorial demonstrated a few of Pixie's community scripts. For more insight into the health of your services, check out the following scripts:

• px/pod shows a CPU flamegraph for the pod to see how your Go/C++/Rust applications are spending their time. To learn more about how use Pixie for application profiling, check out the [Profiling with Flamegraphs](/tutorials/pixie-101/profiler) tutorial.
• px/services shows LET over time for all services in the given namespace, along with a service graph.
• px/service_stats shows LET over time for the given service, along with a service graph and summary of incoming and outgoing traffic.
• px/service_edge_stats shows statistics about the traffic between two services.

Infra HealthDatabase Query Profiling