introduce MilliTimer functions (lyft#131)

This provides an API so that we can move away from microsecond-based
timers.

The StatsD spec suggests using milliseconds, and they're a great fit
for the typical RPCs we are measuring. Most RPCs take about
1-1000ms, which makes it easy for a human to spot and tell the
difference vs. microseconds which would be 1000us-1000000us. We don't
need that amount of fidelity.

Consumers can continue using NewTimer, using microseconds, which might
be appropriate for other use cases like timing functions or tight
in-memory queues. But most new
callers should move toward MilliTimer functions.

tcp_sink: test empty flush (lyft#93)

* tcp_sink: test empty flush

* Run Go 1.18 in CI and update go.mod

Co-authored-by: francisco souza <[email protected]>

Retry failed writes once (lyft#120)

Introduce a retry channel, giving buffers one more chance to be successfully flushed before they're dropped. This is intended to improve handling the case where the underlying connection is intermittently closed, but the destination sink is otherwise generally healthy.

Add list of existing stats to assertion not-found failures (lyft#105)

tags: move tags logic to internal pkg and add ParseTags() and Seriali…

…zeTags() helpers (lyft#103)

Move all tags logic to the internal/tags package and add the SerializeTags() and 
ParseTags() helper functions to the mock package. Moving the tags logic was
required to make it accessible by the mock package. 

This commit also increases a bit and adds some examples to the mock package.

tags: change serializeTags to never modify the tags map (lyft#101)

This changes serializeTags() to not delete entries from the provided
tags map when keys/values are invalid. This prevents a race condition
when users pass the same tags map with invalid entries to
NewCounterWithTags, NewGaugeWithTags, or NewTimerWithTags.

mock: add Counters, Gauges and Timers methods to get a map of all sta…

…ts (lyft#90)

* mock: Sink replace mutex with atomic.Value

This improves Sink performance when called in parallel.

```
benchmark                             old ns/op     new ns/op     delta
BenchmarkFlushCounter-12              50.4          51.9          +2.98%
BenchmarkFlushCounter_Parallel-12     36.5          10.5          -71.23%
BenchmarkFlushTimer-12                52.1          49.8          -4.41%
BenchmarkFlushTimer_Parallel-12       42.0          13.5          -67.86%

benchmark                             old allocs     new allocs     delta
BenchmarkFlushCounter-12              0              0              +0.00%
BenchmarkFlushCounter_Parallel-12     0              0              +0.00%
BenchmarkFlushTimer-12                0              0              +0.00%
BenchmarkFlushTimer_Parallel-12       0              0              +0.00%

benchmark                             old bytes     new bytes     delta
BenchmarkFlushCounter-12              0             0             +0.00%
BenchmarkFlushCounter_Parallel-12     0             0             +0.00%
BenchmarkFlushTimer-12                0             0             +0.00%
BenchmarkFlushTimer_Parallel-12       0             0             +0.00%
```

* mock: add Counters, Gauges and Timers methods to get a map of all stats

This PR also increases test coverage.

subScope: use a slice of tags instead of a map (lyft#97)

* subScope: use a slice of tags instead of a map

This commit changes the subScope to keep a slice of tags instead of a
map. This means that we no longer modify or "own" the map we're passed,
which has been a source of bugs recently (that gostats owns any map
passed to it was poorly, if at all, documented).

This change also improves performance by 2x and reduces memory usage and
GC pressure as each subScope will no longer be holding a map (which is
sparse and the GC must scan).

This change also moves us closer to passing tags via a slice (instead of
a map) which is a change that I intend to make part of the V1 release
(assuming I ever get around to that).

```
benchmark                                old ns/op     new ns/op     delta
BenchmarkStore_ScopeWithTags-12          1321          646           -51.10%
BenchmarkStore_ScopeNoTags-12            583           179           -69.30%

benchmark                                old allocs     new allocs     delta
BenchmarkStore_ScopeWithTags-12          4              3              -25.00%
BenchmarkStore_ScopeNoTags-12            4              2              -50.00%

benchmark                                old bytes     new bytes     delta
BenchmarkStore_ScopeWithTags-12          544           512           -5.88%
BenchmarkStore_ScopeNoTags-12            304           112           -63.16%
```

* Update tags.go

Co-authored-by: Max R <[email protected]>

* stats_test: code review comments

* tags_test: benchmark tagSet.Search

* tags: rename CompareAndSwap => cas and document network sort

* tags: save alloc in replaceChars

Co-authored-by: Max R <[email protected]>

Add setting for disabling logging sink (lyft#94)

* Add setting for disabling logging sink

* Default to current behavior

* Update settings.go

Co-Authored-By: Charlie Vieth <[email protected]>

Co-authored-by: Charlie Vieth <[email protected]>

Add UDP as a sink protocol (lyft#95)

* sink: add UDP support

This commit adds support for UDP statsd sinks with the STATSD_PROTOCOL
environment variable and the WithStatsdProtocol() SinkOption.

* tests: don't run TestHTTPHandler_WrapResponse in parallel

This test is incredibly fast so doesn't need to run in parallel and I'm
trying to reduce the number of goroutines we have in flight at any time
during out tests.