Trace experiments
Execution traces allow for trialing new events on an experimental basis via trace experiments. This document is a guide that explains how you can define your own trace experiments.
Note that if you’re just trying to do some debugging or perform some light
instrumentation, then a trace experiment is way overkill.
Use runtime/trace.Log
instead.
Even if you’re just trying to create a proof-of-concept for a low-frequency
event, runtime/trace.Log
will probably be easier overall if you can make
it work.
Consider a trace experiment if:
- The volume of new trace events will be relatively high, and so the events would benefit from a more compact representation (creating new tables to deduplicate data, taking advantage of the varint representation, etc.).
- It’s not safe to call
runtime/trace.Log
(or its runtime equivalent) in the contexts you want to generate an event (for example, for events about timers).
Defining a new experiment
To define a new experiment, modify internal/trace/tracev2
to define a
new Experiment
enum value.
An experiment consists of two parts: timed events and experimental batches. Timed events are events like any other and follow the same format. They are easier to order and require less work to make use of. Experimental batches are essentially bags of bytes that correspond to an entire trace generation. What they contain and how they’re interpreted is totally up to you, but they’re most often useful for tables that your other events can refer into. For example, the AllocFree experiment uses them to store type information that allocation events can refer to.
Defining new events
- Define your new experiment event types (by convention, experimental events
types start at ID 127, so look for the
const
block defining events starting there). - Describe your new events in
specs
. Use the documentation forSpec
to write your new specs, and check your work by running the tests in theinternal/trace/tracev2
package. If you wish for your event argument to be interpreted in a particular way, follow the naming convention insrc/internal/trace/tracev2/spec.go
. For example, if you intend to emit a string argument, make sure the argument name has the suffixstring
. - Add ordering and validation logic for your new events to
src/internal/trace/order.go
by listing handlers for those events in theorderingDispatch
table. If your events are always emitted in a regular user goroutine context, then the handler should be trivial and just validate the scheduling context to match userGoReqs. If it’s more complicated, see(*ordering).advanceAllocFree
for a slightly more complicated example that handles events from a larger variety of execution environments. If you need to encode a partial ordering, look toward the scheduler events (names beginning withGo
) or just ask someone for help. - Add your new events to the
tracev2Type2Kind
table insrc/internal/trace/event.go
.
Emitting data
Emitting your new events
- Define helper methods on
runtime.traceEventWriter
for emitting your events. - Instrument the runtime with calls to these helper methods.
Make sure to call
traceAcquire
andtraceRelease
around the operation your event represents, otherwise it will not be emitted atomically with that operation completing, resulting in a potentially misleading trace.
Emitting experimental batches
To emit experimental batches, use the runtime.unsafeTraceExpWriter
to
write experimental batches associated with your experiment.
Heed the warnings and make sure that while you write them, the trace
generation cannot advance.
Note that each experiment can only have one distinguishable set of
batches.
Recovering experimental data
Recovering experimental events from the trace
Experimental events will appear in the event stream as an event with the
EventExperimental
Kind
.
Use the Experimental
method to collect the raw data inserted into the
trace.
It’s essentially up to you to interpret the event from here.
I recommend writing a thin wrapper API to present a cleaner interface if you
so desire.
Recovering experimental batches
Parse out all the experimental batches from Sync
events as they come.
These experimental batches are all for the same generation as all the
experimental events up until the next Sync
event.