This makes the makefile's build target have actual dependencies, so that
it only rebuilds any given adapter if that adapter's actual go files
have changed (yes, this is mostly redundant with Go 1.10, but it makes
working on read-only filesystems a bit nicer).
This fixes asynchronous read/write issues to when performing series
discovery by pushing series results onto a channel, instead of trying to
write them directly to a map.
This commit introduces advanced configuration. The rate-interval and
label-prefix flags are removed, and replaced by a configuration file
that allows you to specify series queries and the rules for transforming
those into metrics queries and API resources.
Currently, the adapter uses http.DefaultClient to access Prometheus.
This means that it does support using TLS client certificates, custom
CA certificates, or token-base authentication, which are all common
setups when connecting to Prometheus behind an auth proxy.
This commit adds support for using a separate kubeconfig, or in-cluster
config, to configure auth when connecting to Prometheus.
Fixes #52
[directxman12: cleanups, spelling corrections, and slight refactor]
This allows setting a prefix on the labels used to determine which
resources a series belongs to. The prefix may be set using the
`--label-prefix` flag.
This updates the dependencies to be compatible with Kubernetes 1.8,
including the new custom.metrics.k8s.io/v1beta1 API group and the
corresponding custom-metrics-apiserver changes.
This commit switches to using the boilerplate versions of
a couple different utilities, including the metric info normalization
and the common error types.
Currently, we fetch the discovery information once at the start of
of the adapter, and then never update it. This could prove problematic,
since other API servers might come only after the adapter is started up.
The boilerplate contains a periodically updating RESTMapper that solves
this issue by refreshing the discovery information at an interval (as
specified by a flag), so that we have a chance to fetch new resources.
All errors returned by the provider should be proper Kube API status
errors. This commit cleans up a couple spots that either returned an
invalid error, or were missing an error (such as the case in when no
results matched the requested object).
Previously, if we encountered an error while trying to update our series
list, we'd return an error, aborting the processing of the entire batch.
This could lead to the list of available metrics being severely out of
date. Instead, we simply log an error when we fail to process a metric
name, and skip it.
For vector and range queries, the Prometheus HTTP API takes a timeout
parameter. Previously, we ignored this parameter. Now, we check to see
if the context passed to the query contains a deadline, and if so,
compare that to the current time to calculate an appropriate timeout for
query evaulation.
