This PR modifies how the metrics library handles `Enabled`: previously,
the package `init` decided whether to serve real metrics or just
dummy-types.
This has several drawbacks:
- During pkg init, we need to determine whether metrics are enabled or
not. So we first hacked in a check if certain geth-specific
commandline-flags were enabled. Then we added a similar check for
geth-env-vars. Then we almost added a very elaborate check for
toml-config-file, plus toml parsing.
- Using "real" types and dummy types interchangeably means that
everything is hidden behind interfaces. This has a performance penalty,
and also it just adds a lot of code.
This PR removes the interface stuff, uses concrete types, and allows for
the setting of Enabled to happen later. It is still assumed that
`metrics.Enable()` is invoked early on.
The somewhat 'heavy' operations, such as ticking meters and exp-decay,
now checks the enable-flag to prevent resource leak.
The change may be large, but it's mostly pretty trivial, and from the
last time I gutted the metrics, I ensured that we have fairly good test
coverage.
---------
Co-authored-by: Felix Lange <fjl@twurst.com>
This change includes a lot of things, listed below.
The interfaces have been split up into one write-interface and one read-interface, with `Snapshot` being the gateway from write to read. This simplifies the semantics _a lot_.
Example of splitting up an interface into one readonly 'snapshot' part, and one updatable writeonly part:
```golang
type MeterSnapshot interface {
Count() int64
Rate1() float64
Rate5() float64
Rate15() float64
RateMean() float64
}
// Meters count events to produce exponentially-weighted moving average rates
// at one-, five-, and fifteen-minutes and a mean rate.
type Meter interface {
Mark(int64)
Snapshot() MeterSnapshot
Stop()
}
```
This PR makes the concurrency model clearer. We have actual meters and snapshot of meters. The `meter` is the thing which can be accessed from the registry, and updates can be made to it.
- For all `meters`, (`Gauge`, `Timer` etc), it is assumed that they are accessed by different threads, making updates. Therefore, all `meters` update-methods (`Inc`, `Add`, `Update`, `Clear` etc) need to be concurrency-safe.
- All `meters` have a `Snapshot()` method. This method is _usually_ called from one thread, a backend-exporter. But it's fully possible to have several exporters simultaneously: therefore this method should also be concurrency-safe.
TLDR: `meter`s are accessible via registry, all their methods must be concurrency-safe.
For all `Snapshot`s, it is assumed that an individual exporter-thread has obtained a `meter` from the registry, and called the `Snapshot` method to obtain a readonly snapshot. This snapshot is _not_ guaranteed to be concurrency-safe. There's no need for a snapshot to be concurrency-safe, since exporters should not share snapshots.
Note, though: that by happenstance a lot of the snapshots _are_ concurrency-safe, being unmutable minimal representations of a value. Only the more complex ones are _not_ threadsafe, those that lazily calculate things like `Variance()`, `Mean()`.
Example of how a background exporter typically works, obtaining the snapshot and sequentially accessing the non-threadsafe methods in it:
```golang
ms := metric.Snapshot()
...
fields := map[string]interface{}{
"count": ms.Count(),
"max": ms.Max(),
"mean": ms.Mean(),
"min": ms.Min(),
"stddev": ms.StdDev(),
"variance": ms.Variance(),
```
TLDR: `snapshots` are not guaranteed to be concurrency-safe (but often are).
I also changed the `Sample` type: previously, it iterated the samples fully every time `Mean()`,`Sum()`, `Min()` or `Max()` was invoked. Since we now have readonly base data, we can just iterate it once, in the constructor, and set all four values at once.
The same thing has been done for runtimehistogram.
Back when ResettingTImer was implemented, as part of https://github.com/ethereum/go-ethereum/pull/15910, Anton implemented a `Percentiles` on the new type. However, the method did not conform to the other existing types which also had a `Percentiles`.
1. The existing ones, on input, took `0.5` to mean `50%`. Anton used `50` to mean `50%`.
2. The existing ones returned `float64` outputs, thus interpolating between values. A value-set of `0, 10`, at `50%` would return `5`, whereas Anton's would return either `0` or `10`.
This PR removes the 'new' version, and uses only the 'legacy' percentiles, also for the ResettingTimer type.
The resetting timer snapshot was also defined so that it would expose the internal values. This has been removed, and getters for `Max, Min, Mean` have been added instead.
A lot of types were exported, but do not need to be. This PR unexports quite a lot of them.
metrics: refactor metrics (28035)
This chang creates a GaugeInfo metrics type for registering informational (textual) metrics, e.g. geth version number. It also improves the testing for backend-exporters, and uses a shared subpackage in 'internal' to provide sample datasets and ordered registry.
Implements #21783
---------
Co-authored-by: Martin Holst Swende <martin@swende.se>
This PR adds counter metrics for the CPU system and the Geth process.
Currently the only metrics available for these items are gauges. Gauges are
fine when the consumer scrapes metrics data at the same interval as Geth
produces new values (every 3 seconds), but it is likely that most consumers
will not scrape that often. Intervals of 10, 15, or maybe even 30 seconds
are probably more common.
So the problem is, how does the consumer estimate what the CPU was doing in
between scrapes. With a counter, it's easy ... you just subtract two
successive values and divide by the time to get a nice, accurate average.
But with a gauge, you can't do that. A gauge reading is an instantaneous
picture of what was happening at that moment, but it gives you no idea
about what was going on between scrapes. Taking an average of values is
meaningless.
This changes how we read performance metrics from the Go runtime. Instead
of using runtime.ReadMemStats, we now rely on the API provided by package
runtime/metrics.
runtime/metrics provides more accurate information. For example, the new
interface has better reporting of memory use. In my testing, the reported
value of held memory more accurately reflects the usage reported by the OS.
The semantics of metrics system/memory/allocs and system/memory/frees have
changed to report amounts in bytes. ReadMemStats only reported the count of
allocations in number-of-objects. This is imprecise: 'tiny objects' are not
counted because the runtime allocates them in batches; and certain
improvements in allocation behavior, such as struct size optimizations,
will be less visible when the number of allocs doesn't change.
Changing allocation reports to be in bytes makes it appear in graphs that
lots more is being allocated. I don't think that's a problem because this
metric is primarily interesting for geth developers.
The metric system/memory/pauses has been changed to report statistical
values from the histogram provided by the runtime. Its name in influxdb has
changed from geth.system/memory/pauses.meter to
geth.system/memory/pauses.histogram.
We also have a new histogram metric, system/cpu/schedlatency, reporting the
Go scheduler latency.
