Motivation:
I have a metrics backend where I need to store timer values in
microseconds. The "preferred display unit" models this exactly, but it
seems to have (inadvertently?) omitted the possibility to express
microseconds. Note that among the timer "report" methods there is
already a microseconds option.
Changes:
* change TimeUnit from enum to struct with static members to support non-api breaking future evolution
* add scaleFromNanoseconds to TimeUnit to make it easy to compute the values without switching over
* add tests
Co-authored-by: Chris Burrows <cburrows@gmail.com>
MetricsSystem.bootstrap verifies that the metric system has not been previously
initialized. Otherwise it should fail with a corresponding error message. The
precondition error message includes the name of the currently used factory and
for that accesses self.factory. However, because bootstrap already holds
self.lock as a writer lock, self.factory fails to get it as a reader and
crashes with a less useful precondition error message. This commit ensures that
the correct error message is printed.
motivation: some metrics backend prefer to be given a hint about the preferred display unit (seconds, milliseconds, etc) to drive the ux
changes: add a `preferedUnit` to TimerHandler (and `TimeUnits`) to capture the prefer display unit
motivation: define record() more explicitly that it is a "set the value" operation as it seemed to be confusing when implementing or using the API.
changes: update API docs accordingly
motivaiton: unified format
changes:
* use official docker image 🎉
* fix outstanding formatting issues
* add a call to swiftformat as part of sanity script
* fix sanity script language check debugging statements
Motivation:
As discussed in Issue #28, there are a few preferred lightweight APIs for getting timestamps such as `DispatchTime` but those APIs return `UInt64` types, which need to be handled for overflow.
Modifications:
Added a generic `BinaryInteger` overload method for `recordNanoseconds` that guards against accidental overflows from `UInt64` to `Int64`
Result:
Users now have access to a new method to record nanosecond measurements with `Timer` that guards against overflows.
Motivation:
`Timer.record*` methods are generic to cover all cases of `BinaryInteger` values and converts the passed value into `Int64` before doing time unit conversions and overflow checks.
Modifications:
All methods that accept `BinaryInteger` types and converts to `Int64` now guards against values higher than `Int64.max`
Result:
Users can reliably use the `Timer.record*` APIs with `BinaryInteger` types without concern for unknowing fatal errors.
* doc: #26 clarify which init to use, reformat docs for easier noticing this
* Mark all "we do not expect this API to be used in normal..." as warnings
Motivation:
API documentation contains typos and grammatical errors.
Modifications:
Fix typos and grammatical errors. Update to use code syntax where
appropriate.
Result:
Eliminate typos and grammatical errors found.
motivation: highlight the need to address overflow in metric objects that are integer based
changes:
* add section in the readme about addressing woverflow
* fix timer's unit conversion to address potential overflow
motivation: allow middleware libraries better control of metric object lifecycle to reduce potential memory footprint when using metrics libraries that use some kind of registry or cache (eg polling backends)
changes:
* Introduce metric.destroy() to enable lifecycle management
* Add destroy functions to example impl in readme
* rename Counter::increment "value" parameter to "by"
motivation: nicer API signature
changes: rename Counter::increment and CounterHandler::increment "value" parameter to "by"
* prepare to release
motivation: the sswg voted to adopt the API. this is to prepare to a release
changes:
* rewrite readme
* add API docs
* add utilitiy scripts and docker setup for CI
* adjust linux tests
motivation: it's very beneficial for the protocols to use functions with concrete types because then there's no need for specialisation. At compile time, the compiler can often not know what concrete type will be used so it needs to create a 'generic at runtime' version of the function. If however all functions on the protocol do not use generics, there's no need for a 'generic at runtime' version of the function
changes:
* change CounterHandler::increment to take Int64 instead of <DataType: BinaryInteger>
* change RecorderHandler::record to take Int64 and Double instead of <DataType: BinaryInteger> and <DataType: BinaryFloatingPoint>
* adjust example and test implementation
* adjust docs
motivation: address feedback from community, prepare for moving api to official repo
changes:
* add reset method to counter
* do not force the user facing metric object to implement the mteric handler protocol. this was done for convinience and confuses matters
* adjust tests
* fix a few typos
motivation: after much discussion around logging API, we settled on a different API style, primairly the fact that we will use initializers instead of factories
changes:
* introduce intermediate classes for Counter, Timer and Recorder which are designed to replace the Metrics.makeCounter, Metrics.makeTimer and Metrics.makeRecorder APIs and wrap corresponding CounterHandler, TimerHandler and ReorderHandler coming from the metrics implmentation
* rename Metrics to MetricsSystem
* rename MetricsHandler -> MetricsFactory
* remove Metrics.withCounter, Metrics.withTimer and Metrics.withRecorder syntactic sugar
* rename Metrics.timed with Timer.measure
* make sure metrics system can only be initialized/bootstrapped once per process
* adjust and add tests
* add a bit of docs on key APIs
Konrad `ktoso` Malawski
Moritz Lang
tomer doron
Christian Priebe
Jari (LotU)
Konrad `ktoso` Malawski
tomer doron
Nathan Harris
Nathan Harris
Yim Lee
Konrad `ktoso` Malawski