- noun Perception; understanding.
- noun View; sight.
- intransitive verb To have knowledge or an understanding.
This library provides a set of general-purpose observability tools which can be used to instrument your code in order to better understand it.
In particular, it attempts to satisfy the following goals:
There are a few concepts in ken
to understand in order to use the tools
effectively.
An event is a map of Clojure keys and values which represent a thing that
happened in your code. Events will typically contain a selection of keys from
ken.event
to provide a basic foundation:
{:ken.event/label "the thing",
:ken.event/time #inst "2020-03-27T21:22:27.003Z",
:ken.event/duration 44.362681,
:ken.event/message "Perform some routine activity",
:ken.event/ns my.foo.thing,
,,,}
Here we have an event about some labeled process which started at the given time, lasted 44.3 milliseconds, and had some associated human-friendly metadata. Events may have many other attributes as well, including authentication context, custom identifiers, inputs and outputs, and more.
A context collector in ken
is a source of contextual information used to
enrich the events being observed. Useful data sources could include properties
about the running process such as the build number and deployment environment,
dynamic information about request-specific properties like the authenticated
user, and even local attributes that are breadcrumbs for later deugging.
;; Register a static context which is known at process startup:
(let [build-info {:build-version (System/getProperty "my.build.version")}]
(ken.context/register-static! :build build-info))
;; Context collection happens at the call location, so you can also
;; pull dynamic properties:
(defn user-context
[]
(select-keys my.web.auth/*user-info* [:user/id :user/name :user/email]))
(ken.context/register! :user-info user-context)
Contexts can be removed with unregister!
and the keyword id, or reset
entirely with clear!
.
If you have a service-oriented architecture, a rich source of additional data can be the broadcast context which is transmitted through the whole request graph.
This library adopts the idea of the tap introduced in Clojure 1.10. This is a global queue which may be sent events from anywhere in the code. Note that ken has its own queue, it does not utilize the built-in tap in order to avoid polluting local debugging usage.
In order for those events to be useful, they must be handled by functions which are subscribed to the tap. As an example, we can pretty-print all events to our console for inspection:
(ken.tap/subscribe! :cprint puget.printer/cprint)
Subscribed functions will be called with every event sent to the tap and should
not block for significant periods of time or they may cause event loss.
Subscriptions can be removed with unsubscribe!
or reset entirely with
clear!
.
Finally, ken
uses a standard model for distributed tracing of events. Events
are grouped together under a top-level trace which identifies an entire unit
of work. Each event within this may be a span which represents a subunit of
work covering some duration. Spans may be children of other spans, meaning they
represent more fine-grained bits of work in turn. Linking all the spans in a
trace together forms a tree, sometimes called a "call graph" which represents
the observations collected about the unit of work which was traced.
Using the library will automatically capture and extend the tracing identifiers where needed, which show up in the observed events:
{:ken.event/label "A"
:ken.event/time #inst "2020-03-27T21:22:27.003Z",
:ken.event/duration 44.362681,
:ken.trace/trace-id "bplzs2gajkfcbojkspx7",
:ken.trace/span-id "cuoclyafu4",
,,,}
{:ken.event/label "B"
:ken.event/time #inst "2020-03-27T21:22:27.005Z",
:ken.event/duration 41.805794,
:ken.trace/trace-id "bplzs2gajkfcbojkspx7",
:ken.trace/parent-id "cuoclyafu4",
:ken.trace/span-id "cjatpftw5j",
,,,}
Above are two related spans A
and B
, the second nested inside the first.
Releases are published on Clojars; to use the latest version with Leiningen, add the following to your project dependencies:
Enough theory, how do you actually use this?
(require '[ken.core :as ken])
The most direct way to use the library is to call the observe
macro in your
code in order to send events.
(ken/observe {:ken.event/label "a thing", ::my/key 123})
This will collect the available context and trace data, add it to the event, then send it to the tap for publishing. By default, this returns immediately (non-blocking) but you can specify a timeout in milliseconds if you would like to wait for the event to be accepted. Either way, this returns true if the event was queued and false if it was rejected.
The most common way to generate events is by describing spans which
cover some work happening. You can use the watch
macro for this:
(ken/watch "a thing happening"
(crunch-numbers 2.17 3.14)
(think-heavily "what am I?"))
This will instrument the body of expressions and observes an event at the end
which includes tracing and timing information. The watch
form will return the
value of the final expression. This also works for asynchronous
values, so the following code will
only record the event once the deferred chain completes:
(ken/watch "another thing"
(d/chain
(d/future
(crunch-numbers 8675309))
do-thing-with-result
,,,))
For richer event data, you can specify a map - the string versions above
automatically expand into :ken.event/label
entries:
(ken/watch {:ken.event/label "foo the bar"
::foo 123
::bar 'baz}
(foo-bar! bar))
If you nest calls to watch
, the nested spans will automatically link to the
enclosing watch
as the parent span.
For data you don't necessarily know until you've started doing some work, you
can annotate spans by adding additional properties to the events. When code
is executing inside a watch
, you can use the annotate
, time
, and error
tools:
(ken/watch "a thing"
(try
(when (foo? x)
(ken/annotate {::foo? true}))
(ken/time ::thinking
(think-heavily "what is consciousness?"))
(catch Exception ex
(ken/error ex))))
This would produce a span event labeled "a thing"
with a few potential
additional attributes - a ::foo?
key set to true, a :ken.event/error
key
with the caught exception, and a ::thinking
key holding the number of
milliseconds spent in the think-heavily
call.
In addition to the registered context collector functions, ken comes with a
built-in local context which you can bind anywhere in your code to provide
additional information. Unlike adding keys directly to observe
and watch
calls, these values are propagated to all events inside the block:
;; all events observed inside this block will include the :foo and :bar keys
(ken/with-context {:foo 123
:bar "baz"}
(something-with-foos x y)
,,,)
Sampling is the act of selecting a subset of events from a large collection of events. Not everything needs to be sampled, but if you have high frequency events and most of them are very similar, sampling them can be a good way to reducing your total event volume.
Sampling is controlled by two tracing keys, which can be specified in the
initial ken/watch
or in a later ken/annotate
call.
In order to opt into sampling for a specific span, you can set the
:ken.event/sample-rate
key. This is an integer value n
that will cause, on
average, about 1/n
of the events to be sampled. The rest will be marked to be
discarded by consumers.
When a span has been marked for sampling, it will set the second tracing key
:ken.trace/keep?
on the resulting event. The keep key can have one of three
possible states:
nil
or absent: The span will be kept and forwarded along. This is the
default behavior.false
: The span will be marked to be dropped and the decision will
propagate down to its child spans.true
: The span and its children will be kept. This decision overrides
sampling logic in child spans.The :ken.trace/keep?
key can also be set directly; for example, if you
encounter an error and want to ensure that a span and its (subsequent) children
are recorded, you can use annotate
to set the flag to true.
For additional reading on sampling best practices, see Honeycomb's article on the topic.
NOTE: events which have been "sampled away" are still reported to tap subscribers! It is up to the individual subscribed functions to decide to drop the events or not.
Copyright © 2021 Amperity, Inc.
Distributed under the MIT License.
Can you improve this documentation?Edit on GitHub
cljdoc is a website building & hosting documentation for Clojure/Script libraries
× close