Instrumenting a library

This is an advanced topic. Before reading further, you may want to check whether an instrumentation library for your platform already exists. If not and if you want to take on creating an instrumentation library, first things first; jump on Zipkin Gitter chat channel and let us know. We’ll be extremely happy to help you along the way.

Overview

To instrument a library, you’ll need to understand and create the following elements:

  1. Core data structures - the information that is collected and sent to Zipkin
  2. Trace identifiers - what tags for the information are needed so it can be reassembled in a logical order by Zipkin
    • Generating identifiers - how to generate these IDs and which IDs should be inherited
    • Communicating trace information - additional information that is sent to Zipkin along with the traces and their IDs.
  3. Timestamps and duration - how to record timing information about an operation.

Alright, ready? Here we go.

Core data structures

Core data structures are documented in detail in Thrift comments. Here’s a high-level description to get you started:

Annotation

An Annotation is used to record an occurrence in time. There’s a set of core annotations used to define the beginning and end of an RPC request:

When using message brokers instead of RPCs, the following annotations help clarify the direction of the flow:

Unlike RPC, messaging spans never share a span ID. For example, each consumer of a message is a different child span of the producing span.

Other annotations can be recorded during the request’s lifetime in order to provide further insight. For instance adding an annotation when a server begins and ends an expensive computation may provide insight into how much time is being spent pre and post processing the request versus how much time is spent running the calculation.

BinaryAnnotation

Binary annotations do not have a time component. They are meant to provide extra information about the RPC. For instance when calling an HTTP service, providing the URI of the call will help with later analysis of requests coming into the service. Binary annotations can also be used for exact match search in the Zipkin Api or UI.

Endpoint Annotations and binary annotations have an endpoint associated with them. With two exceptions, this endpoint is associated with the traced process. For example, the service name drop-down in the Zipkin UI corresponds with Annotation.endpoint.serviceName or BinaryAnnotation.endpoint.serviceName. For the sake of usability, the cardinality of Endpoint.serviceName should be bound. For example, it shouldn’t include variables or random numbers.

Span

A set of Annotations and BinaryAnnotations that correspond to a particular RPC. Spans contain identifying information such as traceId, spanId, parentId, and RPC name.

Spans are usually small. For example, the serialized form is often measured in KiB or less. When spans grow beyond orders of KiB, other problems occur, such as hitting limits like Kafka message size (1MiB). Even if you can raise message limits, large spans will increase the cost and decrease the usability of the tracing system. For this reason, be conscious to store data that helps explain system behavior, and don’t store data that doesn’t.

Trace

A set of spans that share a single root span. Traces are built by collecting all Spans that share a traceId. The spans are then arranged in a tree based on spanId and parentId thus providing an overview of the path a request takes through the system.

Trace identifiers

In order to reassemble a set of spans into a full trace three pieces of information are required. Trace identifiers can be 128-bit, but span identifiers within a trace are always 64-bit.

Trace Id

The overall 64 or 128-bit ID of the trace. Every span in a trace shares this ID.

Span Id

The ID for a particular span. This may or may not be the same as the trace id.

Parent Id

This is an optional ID that will only be present on child spans. That is the span without a parent id is considered the root of the trace.

Generating identifiers

Let’s walk through how Spans are identified.

When an incoming request has no trace information attached, we generate a random trace ID and span ID. The span ID can be reused as the lower 64-bits of the trace ID, but it can also be completely different.

If the request already has trace information attached to it, the service should use that information as server receive and server send events are part of the same span as the client send and client receive events

If the service calls out to a downstream service a new span is created as a child of the former span. It is identified by the same trace id, a new span id, and the parent id is set to the span id of the previous span. The new span id should be 64 random bits.

Note This process must be repeated if the service makes multiple downstream calls. That is each subsequent span will have the same trace id and parent id, but a new and different span id.

Communicating trace information

Trace information needs to be passed between upstream and downstream services in order to reassemble a complete trace. Five pieces of information are required:

Check here for the format.

Finagle provides mechanisms for passing this information with HTTP and Thrift requests. Other protocols will need to be augmented with the information for tracing to be effective.

Instrumentation sampling decisions are made at the edge of the system

Downstream services must honour the sampling decision of the upstream system. If there’s no “Sampled” information in the incoming request, the library should make a decision on whether to sample this request, and include the decision in further downstream requests. This simplifies the math when it comes to understanding what’s sampled and what isn’t. It also ensures that a request is either fully traced, or not traced at all, making the sampling policy easier to understand and configure.

Note that the debug flag will force a trace to be sampled, regardless of any sampling rules. The debug flag also applies to storage tier sampling, which is configured on the server side of Zipkin.

HTTP Tracing

HTTP headers are used to pass along trace information.

The B3 portion of the header is so named for the original name of Zipkin: BigBrotherBird.

Ids are encoded as hex strings:

For more information on B3, please see its specification.

Thrift Tracing

Finagle clients and servers negotate whether they can handle extra information in the header of the thrift message when a connection is established. Once negotiated trace data is packed into the front of each thrift message.

Timestamps and duration

Span recording is when timing information or metadata is structured and reported to zipkin. One of the most important parts of this process is appropriately recording timestamps and duration.

Timestamps are microseconds

All Zipkin timestamps are in epoch microseconds (not milliseconds). This value should use the most precise measurement available. For example, clock_gettime or simply multiply epoch milliseconds by 1000. Timestamps fields are stored as 64bit signed integers eventhough negative is invalid.

Microsecond precision primarily supports “local spans”, which are in-process operations. For example, with higher precision, you can tell nuances of what happened before something else.

All timestamps have faults, including clock skew between hosts and the chance of a time service resetting the clock backwards. For this reason, spans should record their duration when possible.

Span duration is also microseconds

While it is possible to get nanosecond-precision timing information, Zipkin uses microsecond granularity. Here are some reasons why:

First, using the same unit as timestamps makes math easier. For example, if you are troubleshooting a span, it is easier to identify with terms in the same unit.

Next, the overhead of recording a span is often variable and can be microseconds or more: suggesting a higher resolution than overhead can be distracting.

Future versions of Zipkin may revisit this topic, but for now, everything is microseconds.

When to set Span.timestamp and duration

Span.timestamp and duration should only be set by the host that started the span.

The simplest logic is generally this:

unless (logging "sr" in an existing span) {
 set Span.timestamp and duration
}

Zipkin merges spans together that share the same trace and span ID. The most common case of this is to merge a span reported by both the client (cs, cr) and the server (sr, ss). For example, the client starts a span, logging “cs” and propagates it via B3 headers, the server continues that span by logging “sr”.

In this case, the client started the span, so it should record Span.timestamp and duration, and those values should match the difference between “cs” and “cr”. The server did not start this span, so it should not set Span.timestamp or duration.

Another common case is when a server starts a root span from an uninstrumented client, such as a web browser. It knows it should start a trace because none was present in B3 headers or similar. Since it started the trace, it should record Span.timestamp and duration on the root span.

Note: When a span is incomplete, you could set Span.timestamp, but not duration as there’s not enough information to do that accurately.

What happens when Span.timestamp and duration are not set?

Span.timestamp and Span.duration are fields added in 2015, 3 years after Zipkin started. Not all libraries log these. When these fields are not set, Zipkin adds them at query time (not collection time); this is not ideal.

The duration query will not work as there’s no data to query. Also, local (in-process) spans aren’t required to have annotations, so they cannot be queried unless their timestamp is set.

When duration isn’t set by instrumentation, Zipkin tries to derive duration at query time, it has to use the problematic method of timestamp math. Ex. if an NTP update happened inside the span, the duration Zipkin caculates will be wrong.

Finally, there’s a desire for many to move to single-host spans. The migration path towards this is to split dual-host RPC spans into two. When instrumentation logs timestamp only for spans it owns, splitting collectors have a heuristic to distinguish a server-initiated root span from a client-initiated, dual-host one.

The bottom-line is that choosing not to record Span.timestamp and duration will result in less accurate data and less functionality. Since it is very easy to record these authoritatively before reporting, all Zipkin instrumentation should do it or ask someone to help them do it.

Message Tracing

Message Tracing is different than RPC tracing because the producer and consumer don’t share span IDs.

In normal RPC tracing, client and server annotations go on the same span. This doesn’t work for messaging because there may be multiple consumers for a given message. The trace context propagated to the consumer is the parent.

Messaging tracing doesn’t have a response path. Only two annotations are used: “ms” and “mr”. Like RPC tracing, it is fine to set Span.timestamp and duration as the producer and each consumer use separate spans.

So, the producer adds “ms” to a span and reports it to zipkin. Then, each consumer creates a child span adding “mr” to it.

Here’s a diagram of Message tracing:

   Producer Tracer                                    Consumer Tracer
+------------------+                               +------------------+
| +--------------+ |     +-----------------+       | +--------------+ |
| | TraceContext |======>| Message Headers |========>| TraceContext | |
| +--------------+ |     +-----------------+       | +--------------+ |
+--------||--------+                               +--------||--------+
   start ||                                                 ||
         \/                                          finish ||
span(context).annotate("ms")                                \/
             .address("ma", broker)          span(context).annotate("mr")
                                                          .address("ma", broker)

Here’s an example of this process using the Brave Tracer:

Producer side:

// Add trace identifiers to the outbound span
tracing.propagation().injector(Message::addHeader)
       .inject(span.context(), message);

producer.send(message);

// start and finish the producer side
span.kind(Span.Kind.PRODUCER)
    .remoteEndpoint(broker.endpoint())
    .start().finish();

// The above will report to zipkin trace identifiers, a "ms" annotation with the
// endpoint of the producer, and a "ma" (Message Address) with the endpoint of
// the broker

Consumer side:

// Parse the span from message headers
TraceContextOrSamplingFlags result =
    tracing.propagation().extractor(Message::getHeader).extract(message);

// Reuse the same span ids by joining that context
span = tracer.newChild(result.context())

// start and finish the consumer side indicating the message arrived.
span.kind(Span.Kind.CONSUMER)
    .remoteEndpoint(broker.endpoint())
    .start().finish();

// The above will report to zipkin trace identifiers, a "mr" annotation with the
// endpoint of the consumer, and a "ma" (Message Address) with the endpoint of
// the broker.

Many consumers act in bulk, receiving many messages at the same time. It may be helpful to inject each consumer span’s trace context into its corresponding message headers. This allows a processor to create a child later, at the right place in the trace tree.

Here’s an example of doing this with Kafka’s poll api:

public ConsumerRecords<K, V> poll(long timeout) {
  ConsumerRecords<K, V> records = delegate.poll(timeout);
  for (ConsumerRecord<K, V> record : records) {
    handleConsumed(record);
  }
  return records;
}

void handleConsumed(ConsumerRecord record) {
  // notifies zipkin the record arrived
  Span span = startAndFinishConsumerSpan(record);
  // allows a processor to see the parent ID (the consumer trace context)
  injector.inject(span.context(), record.headers());
}