[Metric] use Float64Counter to get wrong metric data

[gunsluo]

Describe the bug
use Float64Counter to get the wrong metric data.

the code

meterPusher = push.New(
			simple.NewWithExactDistribution(),
			exporter,
			push.WithPeriod(30*time.Second),
			//push.WithTimeout(10*time.Second),
		)
meterProvider = meterPusher.Provider()
meter:= meterProvider.Meter("name")

accountReadCounter : =meter.NewFloat64Counter("account.read", metric.WithDescription("record number of reading account"));

....

accountReadCounter.Add(ctx, 1)

the counter metric's value should not be reduced.

What did you expect to see?
the value of the counter keeps increasing by time.

What version did you use?
Version:
collector 0.5.0
otel: 0.7.0

What config did you use?
collector Config:

receivers:
  otlp:
    protocols:
      grpc:
        endpoint: "0.0.0.0:55679"

processors:
  batch:
  queued_retry:

extensions:
  health_check: {}

exporters:
  jaeger:
    endpoint: "jaeger:14250"
    insecure: true
  prometheus:
    endpoint: "0.0.0.0:8889"
    namespace: example
    const_labels:
      label1: test
  logging:
      loglevel: debug

service:
  extensions: [health_check]
  pipelines:
    traces:
      receivers: [otlp]
      processors: [batch, queued_retry]
      exporters: [jaeger]

    metrics:
      receivers: [otlp]
      exporters: [prometheus, logging]

Environment
OS: macOS
Compiler: go version go1.14.2 darwin/amd64

Additional context
it works fine on version 0.3.0(collector) and otel(0.6.0).

[gunsluo]

Hi, @nilebox

I found that SDK reset the value of the counter when record's updateCount is not collectedCount, there is no any modification of collectedCount in code. so the value of the counter is reset.

https://github.com/open-telemetry/opentelemetry-go/blob/master/sdk/metric/sdk.go#L376

opentelemetry-go/sdk/metric/sdk.go

Line 437 in 27e892a

func (m *Accumulator) checkpointRecord(r *record) int {

is it a bug?

[Aneurysm9]

@jmacd is this related to #887/#903?

[jmacd]

I am a bit confused. From the report, it looks more like version 0.6 is still being used. The 0.7 release added metrics processor logic to sum the individual counts, and the problems in #887 and #903 would have produced different odd behavior. Is there any chance this is a 0.6 otel library with a 0.4 collector? Otherwise, it's an unusual report and I would like to continue investigating.

[jmacd]

This is a case of open-telemetry/opentelemetry-collector#1255. The OTLP to Prometheus functionality is broken in the 0.4 collector release: we will disable the OTLP receiver in the 0.5 release and get this fixed in 0.6 I hope.

[jmacd]

Closing this as it's covered in collector issue 1255.