B3 Propagation is a specification for the header "b3" and those that start with "x-b3-". These headers are used for trace context propagation across service boundaries.
This specification elaborates identifiers used to place an operation in a trace tree. These attributes are propagated in-process, and eventually downstream (often via http headers), to ensure all activity originating from the same root are collected together. A sampling decision is made at the root of the trace, and this indicates if trace details should be collected and reported to the tracing system (usually Zipkin) or not.
The most common propagation use case is to copy a trace context from a client sending an RPC request to a server receiving it.
In this case, the same span ID is used, which means that both the client and server side of an operation end up in the same node in the trace tree.
Here's an example flow using multiple header encoding, assuming an HTTP request carries the propagated trace:
Client Tracer Server Tracer
┌───────────────────────┐ ┌───────────────────────┐
│ │ │ │
│ TraceContext │ Http Request Headers │ TraceContext │
│ ┌───────────────────┐ │ ┌───────────────────┐ │ ┌───────────────────┐ │
│ │ TraceId │ │ │ X-B3-TraceId │ │ │ TraceId │ │
│ │ │ │ │ │ │ │ │ │
│ │ ParentSpanId │ │ Inject │ X-B3-ParentSpanId │ Extract │ │ ParentSpanId │ │
│ │ ├─┼────────>│ ├─────────┼>│ │ │
│ │ SpanId │ │ │ X-B3-SpanId │ │ │ SpanId │ │
│ │ │ │ │ │ │ │ │ │
│ │ Sampling decision │ │ │ X-B3-Sampled │ │ │ Sampling decision │ │
│ └───────────────────┘ │ └───────────────────┘ │ └───────────────────┘ │
│ │ │ │
└───────────────────────┘ └───────────────────────┘
Trace identifiers are often sent with a sampling decision. However, it is valid and common practice to send a sampling decision independently. Here's an example of a proxy forbidding tracing of a /health
endpoint. The diagram shows the receiver making a NoOp trace context to ensure minimal overhead.
Server Tracer
┌───────────────────────┐
Health check request │ │
┌───────────────────┐ │ TraceContext │
│ GET /health │ Extract │ ┌───────────────────┐ │
│ X-B3-Sampled: 0 ├─────────┼>│ NoOp │ │
└───────────────────┘ │ └───────────────────┘ │
└───────────────────────┘
Trace identifiers are 64 or 128-bit, but all span identifiers within a trace are 64-bit. All identifiers are opaque.
Identifiers are almost always sent with Sampling state, but they can be sent alone to implement a "Defer" decision described below.
The TraceId is 64 or 128-bit in length and indicates the overall ID of the trace. Every span in a trace shares this ID.
The SpanId is 64-bit in length and indicates the position of the current operation in the trace tree. The value should not be interpreted: it may or may not be derived from the value of the TraceId.
The ParentSpanId is 64-bit in length and indicates the position of the parent operation in the trace tree. When the span is the root of the trace tree, there is no ParentSpanId.
Sampling is a mechanism to reduce the volume of data that ends up in the tracing system. In B3, sampling applies consistently per-trace: once the sampling decision is made, the same value should be consistently sent downstream. This means you will see all spans sharing a trace ID or none.
Here are the valid sampling states. Note they all applied to the trace ID, not the span ID:
- Defer: aka I don't know yet!
- Defer is used when trace identifiers are set by a proxy, but that proxy doesn't send data to Zipkin. The most common use case for defering a decision is pre-provisioning trace identifiers.
- In all known encodings, defer is the absence of sampling state.
- Deny: aka don't sample or don't record
- Deny is used to achieve a probabilistic rate or to prevent certain paths (such as health checks) from generating traces. Where possible, instrumentation should optimize deny such that less overhead occurs.
- Accept: aka sample or record
- Accept is used to achieve a probabilistic rate or to ensure certain paths (such as low-traffic endpoints) are always traced. When a trace is accepted, spans should be reported to zipkin except in overload scenarios.
- Debug: aka force trace
- Debug is a production troubleshooting aid used in tools like curl or chrome debug. Debug is an emphasized accept decision that implies accept, additionally reporting
Span.debug = true
for each span in the trace.
- Debug is a production troubleshooting aid used in tools like curl or chrome debug. Debug is an emphasized accept decision that implies accept, additionally reporting
The most common use of sampling is probablistic: eg, accept 0.01% of traces and deny the rest. Debug is the least common use case.
Sampling state is almost always sent with Identifiers, but it can be sent alone (a predefined decision). This is used for a few use cases:
- Deny: Using a constant entry for a deny decision is more efficient than generating identifiers. Libraries that do this cannot do external ID correlation, though, for example, log correlation.
- Accept: Some proxies send a constant entry for an accept decision as they want a specific path to be traced, but don't want to interfere with ID generation.
- Debug: Some support guides send debug without IDs as it is easier to do in curl.
Below are a number of predefined encodings. The oldest encoding is multiple HTTP headers (2012). Some application notes have been made since, as well a compact, single-header encoding. A binary encoding may be defined as well.
Most users do not need to know the encodings listed here as they are built-in to libraries, frameworks and proxies. While this specification aims to be complete, it will not elaborate all use cases and practice. Please contact the Zipkin gitter if you are confused about something.
We recommend re-using existing encodings over developing new ones as there are usually subtleties that even experienced folks overlook. For example, one proxy accidentally chose to use a hyphen to substitute for an absent parent span ID. This caused broken traces, and in some extreme cases, crashed requests. Some types of software have multiple month or longer release cycles, which adds tension to add "spaghetti" to clean it up. Identifying and answering questions about such problems add support load unfairly to others as they weren't consulted. Even ruling out format mismatches, misunderstanding of the feature set can cause expensive maintenance later.
The best way to avoid costly mistakes is to re-use an encoding. The second-best way is reading this spec completely when designing a bespoke encoding, and getting feedback from the authors on your work. If you are designing a new encoding, please contact the Zipkin gitter so that you have the best chance to avoid mistakes others made in the past.
There are two encodings of B3: Single Header and Multiple Header. Multiple header encoding uses an X-B3-
prefixed header per item in the trace context. Single header delimits the context into into a single entry named b3
. The single-header variant takes precedence over the multiple header one when extracting fields.
B3 attributes are most commonly propagated as multiple http headers. All B3 headers follows the convention of X-B3-${name}
with special-casing for flags. When extracting state headers, the first value wins.
Note: Http headers are case-insensitive, but sometimes this encoding is used for other transports. When encoding in case-sensitive transports, prefer lowercase keys or the single header header encoding which is explicitly lowercase.
As mentioned earlier, identifiers can be sent with or without sampling state and visa versa. It is important to understand the relationships between these headers. For example, X-B3-TraceId
and X-B3-SpanId
can be sent alone, or with a sampling header X-B3-Sampled
. It is also valid to send sampling state independently, such as a deny decision X-B3-Sampled: 0
.
Note: Two headers, X-B3-Sampled
and X-B3-ParentSpanId
can be absent. Absent means the header doesn't exist. An empty header (ex. X-B3-Sampled:
) or an arbitrary nonsense value (ex X-B3-ParentSpanId: -
) are examples of malformed data.
The X-B3-TraceId
header is encoded as 32 or 16 lower-hex characters. For example, a 128-bit TraceId header might look like: X-B3-TraceId: 463ac35c9f6413ad48485a3953bb6124
. Unless propagating only the Sampling State, the X-B3-TraceId
header is required.
The X-B3-SpanId
header is encoded as 16 lower-hex characters. For example, a SpanId header might look like: X-B3-SpanId: a2fb4a1d1a96d312
. Unless propagating only the Sampling State, the X-B3-SpanId
header is required.
The X-B3-ParentSpanId
header may be present on a child span and must be absent on the root span. It is encoded as 16 lower-hex characters. For example, a ParentSpanId header might look like: X-B3-ParentSpanId: 0020000000000001
An accept sampling decision is encoded as X-B3-Sampled: 1
and a deny as X-B3-Sampled: 0
. Absent means defer the decision to the receiver of this header. For example, a Sampled header might look like: X-B3-Sampled: 1
.
Note: Before this specification was written, some tracers propagated X-B3-Sampled
as true
or false
as opposed to 1
or 0
. While you shouldn't encode X-B3-Sampled
as true
or false
, a lenient implementation may accept them.
Debug is encoded as X-B3-Flags: 1
. Absent or any other value can be ignored. Debug implies an accept decision, so don't also send the X-B3-Sampled
header.
A single header named b3
standardized in late 2018 for use in JMS and w3c tracestate
. Design and rationale are captured here. Check or update our status page for adoption status.
In simplest terms b3
maps propagation fields into a hyphen delimited string.
b3={TraceId}-{SpanId}-{SamplingState}-{ParentSpanId}
, where the last two fields are optional.
For example, the following state encoded in multiple headers:
X-B3-TraceId: 80f198ee56343ba864fe8b2a57d3eff7
X-B3-ParentSpanId: 05e3ac9a4f6e3b90
X-B3-SpanId: e457b5a2e4d86bd1
X-B3-Sampled: 1
Becomes one b3
header, for example:
b3: 80f198ee56343ba864fe8b2a57d3eff7-e457b5a2e4d86bd1-1-05e3ac9a4f6e3b90
Note: When only propagating a sampling decision, the header is still named b3
, but only contains the sampling state.
A deny decision encodes as:
b3: 0
The first position of the b3
value is the 32 or 16 lower-hex character TraceId. Unless propagating
only the Sampling State, the TraceId is required.
For example, a TraceId of 80f198ee56343ba864fe8b2a57d3eff7
encodes in the left-most position:
b3: 80f198ee56343ba864fe8b2a57d3eff7-e457b5a2e4d86bd1-1
The second position of the b3
value is the 16 lower-hex character SpanId. Unless propagating only
the Sampling State, the SpanId is required.
For example, a SpanId of e457b5a2e4d86bd1
encodes after the left-most hyphen:
b3: 80f198ee56343ba864fe8b2a57d3eff7-e457b5a2e4d86bd1-1
When the sampling state isn't Defer, it is encoded as a single hex character in the third position
of the b3
value. Sampling state can also be sent on its own (with no identifiers).
Sampling State is encoded as a single hex character for all states except Defer. Defer is absence of the sampling field.
- Accept:
1
- Deny:
0
- Debug:
d
For example, a debug trace could look like this
b3: 80f198ee56343ba864fe8b2a57d3eff7-e457b5a2e4d86bd1-d
A deny decision may omit identifiers and look like this:
b3: 0
When present, ParentSpanId is the 16 lower-hex characters in the final position of b3
.
For example, a ParentSpanId of 05e3ac9a4f6e3b90
encodes in the right-most position:
b3: 80f198ee56343ba864fe8b2a57d3eff7-e457b5a2e4d86bd1-1-05e3ac9a4f6e3b90
B3 attributes can also be propagated as ASCII headers in the Custom Metadata of a request. The encoding is exactly the same as Http headers, except the names are explicitly or implicitly down-cased.
For example, when using multiple headers, X-B3-ParentSpanId: 0020000000000001
encodes
as an ASCII header x-b3-parentspanid
with the same value.
JMS (Java Message Service) has constraints that disallow X-B3-
prefixed headers.
As such, only the single header format b3
should be used with JMS. As messaging spans
never share a SpanId, it is encouraged to omit the ParentSpanId field when propagating.
Refer to our design documentation for details.
In B3, the trace context is extracted from incoming headers. Timing and
metadata for the client and server side of an operation are recorded with
the same context. The ParentSpanId
is the ID of the operation that caused
the current RPC. For example, it could be the ID of another server request
or a scheduled job. ParentSpanId
is propagated so that when data is reported
to Zipkin, it can be placed in the correct spot in the trace tree.
Here's an example of a B3 library extracting a trace context from incoming http request headers:
┌───────────────────┐
Incoming Headers │ TraceContext │
┌───────────────────┐ │ ┌───────────────┐ │
│ X-B3-TraceId │──────┼─┼> TraceId │ │
│ │ │ │ │ │
│ X-B3-ParentSpanId │──────┼─┼> ParentSpanId │ │
│ │ │ │ │ │
│ X-B3-SpanId │──────┼─┼> SpanId │ │
│ │ │ └───────────────┘ │
└───────────────────┘ └───────────────────┘
Some propagation formats look similar to B3, but don't propagate a field named
parent. Instead, they propagate a span ID field which serves the same purpose
as ParentSpanId
. Unlike B3, these systems use a different span ID for the
client and server side of an RPC. When a server reads headers like this, it is
expected to provision a new span ID for itself, and use the one it extracted as
its parent.
Here's an example of an alternate library composing a trace context with incoming http request headers and an ID generator:
┌───────────────────┐
Incoming Headers │ TraceContext │
┌───────────────────┐ │ ┌───────────────┐ │
│ XXXX-TraceId │──────┼─┼> TraceId │ │
│ │ │ │ │ │
│ XXXX-SpanId │──────┼─┼> ParentSpanId │ │
└───────────────────┘ │ │ │ │ ┌──────────────┐
│ │ SpanId <┼─┼──────│ ID Generator │
│ └───────────────┘ │ └──────────────┘
└───────────────────┘
In both B3 and the above example, incoming headers contain the parent's span ID, and three IDs (trace, parent, span) end up in the trace context. The difference is that B3 uses the same span ID for the client and server side of an RPC, where the latter does not.
It may not be obvious why you'd send a reject sampling decision to the next hop. Imagine a service decides not to trace an operation and makes 2 out-going calls, and these branched out further. If reject ("don't trace") isn't propagated, the system might receive only parts of the operation, confusing users.
Another reason to reject sampling is to prevent overwhelming the collector with high throughput. If the reject-sampling decision is not propagated then there's no way to tell recipients of outbound calls that they should not send span info to the collector.
While valid to propagate only a reject sampling decision (X-B3-Sampled: 0
),
if trace identifiers are established, they should be propagated, too. The
tracing system is often not the only consumer of trace identifiers. Services may
still want to do things with the tracing info even if they aren't sending span
data to the collector, e.g. tag logs with trace IDs.
Deferring the sampling decision is special-case. The only known use-cases are the following:
- Debug trace: A debug trace is implicitly an accepted trace
- Externally provisioned IDs: When you want to control IDs, but not sampling policy
Unless it is a debug trace, leaving out a sampling decision is typically ID correlation. For example, someone re-uses a global identifier from another system, or correlating in logs. In these cases, the caller knows the ID they want, but allows the next hop to decide if it will be traced or not. The caller should not report a span to the tracing system using this ID unless they propagate an accept sampling decision.
The first tracer to use B3 was Finagle. In Finagle, thrift-rpc was a dominant protocol, and a fixed-width binary encoding held B3 attributes. This binary encoding used a bit field to encode the sampling decision and the debug flag.
Http encoding chose to use a separate header for the sampling decision, but
kept the flags field for debug. In hind sight, it would have made more sense to
use a separate header such as X-B3-Debug: 1
as no other flags were added and
no implementation treats X-B3-Flags
as a bit field. That said, renaming the
field would cause more damage than good, so it was left alone.
When you see B3 headers of any kind, you cannot assume the out-of-band data is in Zipkin format, or even that data is sent to a tracing system. For example, some embed trace data into log lines correlated with the same IDs in the headers. Others use B3 headers for dramatically different trace formats.
That said, it is typically the case that data is sent out-of-band, and outgoing fields match some fields in that data. For example, in Zipkin format, all fields except Accept and Deny flags are reported. That's because Accept is implied and Deny would not result in reported data.
It is possible that a proxy, such as Envoy, accepts trace data correlated with B3 headers sent by the application. Depending on configuration, that proxy could start a trace for the act of sending trace data, amplifying traffic to your tracing system with low or negative value traces. While most sites will not trace data sent to a tracing system, there are generally 2 ways to address this proxy setup:
- Configure the proxy to not trace your trace reporting endpont (Ex in Zipkin,
POST /api/v2/spans
) - Send
b3: 0
header (an explicit Deny), to prevent tracing.
We recommend having your trace reporting library send b3: 0
. This cements a
cost to everyone, only to help those proxying their trace data. However, it is
becoming common in mesh environments to proxy everything, and in reality there is
limited ability for end users to change the sampling policy of proxies.
Note: we don't suggest sending also x-b3-sampled: 0
as the proxy problem is
recent and largely contained to Envoy, which already supports B3 single format.
Sending a larger second header would increase the pain to others who don't have
this problem anyway. Put another way, the few proxies missing B3 single header
support should update instead of burdening more people from lack thereof.