采样

使用分布式跟踪，您可以观察请求在分布式系统中从一个服务移动到另一个服务的过程。 它非常实用，原因有很多，比如理解您的服务连接和诊断延迟问题，还有许多其他好处。

However, if the majority of all your requests are successful 200s and finish without unacceptable latency or errors, do you really need all that data? Here’s the thing—you don’t always need a ton of data to find the right insights. You just need the right sampling of data.

The idea behind sampling is to control the spans you send to your observability backend, resulting in lower ingest costs. Different organizations will have their own reasons for not just why they want to sample, but also what they want to sample. You might want to customize your sampling strategy to:

• Manage costs: If you have a high volume of telemetry, you risk incurring heavy charges from a telemetry backend vendor or cloud provider to export and store every span.
• Focus on interesting traces: For example, your frontend team may only want to see traces with specific user attributes.
• Filter out noise: For example, you may want to filter out health checks.

术语

It's important to use consistent terminology when discussing sampling. A trace or span is considered "sampled" or "not sampled":

• Sampled: A trace or span is processed and exported. Because it is chosen by the sampler as a representative of the population, it is considered "sampled".
• Not sampled: A trace or span is not processed or exported. Because it is not chosen by the sampler, it is considered "not sampled".

Sometimes, the definitions of these terms get mixed up. You may find someone state that they are "sampling out data" or that data not processed or exported is considered "sampled". These are incorrect statements.

头抽样

Head sampling is a sampling technique used to make a sampling decision as early as possible. A decision to sample or drop a span or trace is not made by inspecting the trace as a whole.

For example, the most common form of head sampling is Consistent Probability Sampling. It may also be referred to as Deterministic Sampling. In this case, a sampling decision is made based on the trace ID and a desired percentage of traces to sample. This ensures that whole traces are sampled - no missing spans - at a consistent rate, such as 5% of all traces.

The upsides to head sampling are:

• Easy to understand
• Easy to configure
• Efficient
• Can be done at any point in the trace collection pipeline

The primary downside to head sampling is that it is not possible make a sampling decision based on data in the entire trace. This means that head sampling is effective as a blunt instrument, but is wholly insufficient for sampling strategies that must take whole-system information into account. For example, it is not possible to use head sampling to ensure that all traces with an error within them are sampled. For this, you need Tail Sampling.

尾巴抽样

Tail sampling is where the decision to sample a trace takes place by considering all or most of the spans within the trace. Tail Sampling gives you the option to sample your traces based on specific criteria derived from different parts of a trace, which isn’t an option with Head Sampling.

Some examples of how you can use Tail Sampling include:

• Always sampling traces that contain an error
• Sampling traces based on overall latency
• Sampling traces based on the presence or value of specific attributes on one or more spans in a trace; for example, sampling more traces originating from a newly deployed service
• Applying different sampling rates to traces based on certain criteria

As you can see, tail sampling allows for a much higher degree of sophistication. For larger systems that must sample telemetry, it is almost always necessary to use Tail Sampling to balance data volume with usefulness of that data.

There are three primary downsides to tail sampling today:

• Tail sampling can be difficult to implement. Depending on the kind of sampling techniques available to you, it is not always a "set and forget" kind of thing. As your systems change, so too will your sampling strategies. For a large and sophisticated distributed system, rules that implement sampling strategies can also be large and sophisticated.
• Tail sampling can be difficult to operate. The component(s) that implement tail sampling must be stateful systems that can accept and store a large amount of data. Depending on traffic patterns, this can require dozens or even hundreds of nodes that all utilize resources differently. Furthermore, a tail sampler may need to "fall back" to less computationally-intensive sampling techniques if it is unable to keep up with the volume of data it is receiving. Because of these factors, it is critical to monitor tail sampling components to ensure that they have the resources they need to make the correct sampling decisions.
• Tail samplers often end up being in the domain of vendor-specific technology today. If you're using a paid vendor for Observability, the most effective tail sampling options available to you may be limited to what the vendor offers.

Finally, for some systems, tail sampling may be used in conjunction with Head Sampling. For example, a set of services that produce an extremely high volume of trace data may first use head sampling to only sample a small percentage of traces, and then later in the telemetry pipeline use tail sampling to make more sophisticated sampling decisions before exporting to a backend. This is often done in the interest of protecting the telemetry pipeline from being overloaded.