Skip to content

Prometheus instrumentation library for gRPC powered Go applications.

License

Notifications You must be signed in to change notification settings

piotrkowalczuk/promgrpc

Repository files navigation

promgrpc GoDoc

Package promgrpc is an instrumentation package that allows capturing metrics of your gRPC based services, both the server and the client side. The main goal of version 4 was to make it modular without sacrificing the simplicity of use.

It is still possible to integrate the package in just a few lines. However, if necessary, metrics can be added, removed or modified freely.

Design

The package does not introduce any new concepts to an already complicated environment. Instead, it focuses on providing implementations of interfaces exported by gRPC and Prometheus libraries.

It causes no side effects nor has global state. Instead, it comes with handy one-liners to reduce integration overhead.

The package achieved high modularity by using Inversion of Control. We can define three layers of abstraction, where each is configurable or if necessary replaceable.

Collectors

Collectors serve one purpose, storing metrics. These are types well-known from Prometheus ecosystem, like counters, gauges, histograms or summaries. This package comes with a set of predefined functions that create a specific instances for each use case. For example:

func NewRequestsTotalCounterVec(Subsystem, ...CollectorOption) *prometheus.CounterVec

StatsHandlers

Level higher consist of stats handlers. This layer is responsible for metrics collection. It is aware of a collector and knows how to use it to record event occurrences. Each implementation satisfies stats.Handler and prometheus.Collector interface and knows how to monitor a single dimension, e.g. a total number of received/sent requests:

 func NewClientRequestsTotalStatsHandler(*prometheus.GaugeVec, ...StatsHandlerOption) *ClientRequestsTotalStatsHandler
 func NewServerRequestsTotalStatsHandler(*prometheus.GaugeVec, ...StatsHandlerOption) *ServerRequestsTotalStatsHandler

Coordinators

Above all, there is a coordinator. StatsHandler combines multiple stats handlers into a single instance.

Metrics

The package comes with eighteen predefined metrics — nine for server side and nine for client side:

 grpc_client_connections
 grpc_client_message_received_size_histogram_bytes
 grpc_client_message_sent_size_histogram_bytes
 grpc_client_messages_received_total
 grpc_client_messages_sent_total
 grpc_client_request_duration_histogram_seconds
 grpc_client_requests_in_flight
 grpc_client_requests_sent_total
 grpc_client_responses_received_total
 grpc_server_connections
 grpc_server_message_received_size_histogram_bytes
 grpc_server_message_sent_size_histogram_bytes
 grpc_server_messages_received_total
 grpc_server_messages_sent_total
 grpc_server_request_duration_histogram_seconds
 grpc_server_requests_in_flight
 grpc_server_requests_received_total
 grpc_server_responses_sent_total

Configuration

The package does not require any configuration whatsoever but makes it possible. It is beneficial for different reasons.

Having all metrics enabled could not be desirable. Some, like histograms, can create significant overhead on the producer side. If performance is critical, it advisable to reduce the set of metrics. To do that, implement a custom version of coordinator constructor, ClientStatsHandler and/or ServerStatsHandler.

Another good reason to change default settings is backward compatibility. Migration of Grafana dashboards is not an easy nor quick task. If the discrepancy is small and, e.g. the only necessary adjustment is changing the namespace, it is achievable by passing CollectorWithNamespace to a collector constructor. It is the same well-known pattern from the gRPC package, with some enhancements. What makes it different is that both StatsHandlerOption and CollectorOption have a shareable variant, called ShareableCollectorOption and ShareableStatsHandlerOption respectively. Thanks to that, it is possible to pass options related to stats handlers and collectors to coordinator constructors. Constructors take care of moving options to the correct receivers.

Mixing both strategies described above will give even greater freedom. However, if that is even not enough, it is possible to reimplement an entire stack for a given metric or metrics.

Version comparison

version package management api grpc client prometheus client docs
v1.0.2 glide interceptors ^1.0.0 ~0.8.0
v2.0.2 modules interceptors + stats.Handler 1.13.0 0.8.0 GoDoc
v3.2.2 modules interceptors + stats.Handler 1.13.0 0.8.0 GoDoc
v4.1.4 modules stats.Handler 1.66.0 1.20.2 GoDoc

Example

import "github.com/piotrkowalczuk/promgrpc/v4"

// ...

reg := prometheus.NewRegistry()

ssh := promgrpc.ServerStatsHandler(
    promgrpc.CollectorWithNamespace("example"),
    promgrpc.CollectorWithConstLabels(prometheus.Labels{"service": "foo"}),
)
csh := promgrpc.ClientStatsHandler(
    promgrpc.CollectorWithConstLabels(prometheus.Labels{"service": "bar"}),
)

srv := grpc.NewServer(grpc.StatsHandler(ssh))
imp := newDemoServer()

test.RegisterTestServiceServer(srv, imp)
reg.MustRegister(ssh)
reg.MustRegister(csh)

Development

make setup