khenaidoo | ab1f7bd | 2019-11-14 14:00:27 -0500 | [diff] [blame] | 1 | // Copyright 2016 The Prometheus Authors |
| 2 | // Licensed under the Apache License, Version 2.0 (the "License"); |
| 3 | // you may not use this file except in compliance with the License. |
| 4 | // You may obtain a copy of the License at |
| 5 | // |
| 6 | // http://www.apache.org/licenses/LICENSE-2.0 |
| 7 | // |
| 8 | // Unless required by applicable law or agreed to in writing, software |
| 9 | // distributed under the License is distributed on an "AS IS" BASIS, |
| 10 | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 11 | // See the License for the specific language governing permissions and |
| 12 | // limitations under the License. |
| 13 | |
| 14 | // Package promhttp provides tooling around HTTP servers and clients. |
| 15 | // |
| 16 | // First, the package allows the creation of http.Handler instances to expose |
| 17 | // Prometheus metrics via HTTP. promhttp.Handler acts on the |
| 18 | // prometheus.DefaultGatherer. With HandlerFor, you can create a handler for a |
| 19 | // custom registry or anything that implements the Gatherer interface. It also |
| 20 | // allows the creation of handlers that act differently on errors or allow to |
| 21 | // log errors. |
| 22 | // |
| 23 | // Second, the package provides tooling to instrument instances of http.Handler |
| 24 | // via middleware. Middleware wrappers follow the naming scheme |
| 25 | // InstrumentHandlerX, where X describes the intended use of the middleware. |
| 26 | // See each function's doc comment for specific details. |
| 27 | // |
| 28 | // Finally, the package allows for an http.RoundTripper to be instrumented via |
| 29 | // middleware. Middleware wrappers follow the naming scheme |
| 30 | // InstrumentRoundTripperX, where X describes the intended use of the |
| 31 | // middleware. See each function's doc comment for specific details. |
| 32 | package promhttp |
| 33 | |
| 34 | import ( |
| 35 | "compress/gzip" |
| 36 | "fmt" |
| 37 | "io" |
| 38 | "net/http" |
| 39 | "strings" |
| 40 | "sync" |
| 41 | "time" |
| 42 | |
| 43 | "github.com/prometheus/common/expfmt" |
| 44 | |
| 45 | "github.com/prometheus/client_golang/prometheus" |
| 46 | ) |
| 47 | |
| 48 | const ( |
| 49 | contentTypeHeader = "Content-Type" |
| 50 | contentEncodingHeader = "Content-Encoding" |
| 51 | acceptEncodingHeader = "Accept-Encoding" |
| 52 | ) |
| 53 | |
| 54 | var gzipPool = sync.Pool{ |
| 55 | New: func() interface{} { |
| 56 | return gzip.NewWriter(nil) |
| 57 | }, |
| 58 | } |
| 59 | |
| 60 | // Handler returns an http.Handler for the prometheus.DefaultGatherer, using |
| 61 | // default HandlerOpts, i.e. it reports the first error as an HTTP error, it has |
| 62 | // no error logging, and it applies compression if requested by the client. |
| 63 | // |
| 64 | // The returned http.Handler is already instrumented using the |
| 65 | // InstrumentMetricHandler function and the prometheus.DefaultRegisterer. If you |
| 66 | // create multiple http.Handlers by separate calls of the Handler function, the |
| 67 | // metrics used for instrumentation will be shared between them, providing |
| 68 | // global scrape counts. |
| 69 | // |
| 70 | // This function is meant to cover the bulk of basic use cases. If you are doing |
| 71 | // anything that requires more customization (including using a non-default |
| 72 | // Gatherer, different instrumentation, and non-default HandlerOpts), use the |
| 73 | // HandlerFor function. See there for details. |
| 74 | func Handler() http.Handler { |
| 75 | return InstrumentMetricHandler( |
| 76 | prometheus.DefaultRegisterer, HandlerFor(prometheus.DefaultGatherer, HandlerOpts{}), |
| 77 | ) |
| 78 | } |
| 79 | |
| 80 | // HandlerFor returns an uninstrumented http.Handler for the provided |
| 81 | // Gatherer. The behavior of the Handler is defined by the provided |
| 82 | // HandlerOpts. Thus, HandlerFor is useful to create http.Handlers for custom |
| 83 | // Gatherers, with non-default HandlerOpts, and/or with custom (or no) |
| 84 | // instrumentation. Use the InstrumentMetricHandler function to apply the same |
| 85 | // kind of instrumentation as it is used by the Handler function. |
| 86 | func HandlerFor(reg prometheus.Gatherer, opts HandlerOpts) http.Handler { |
| 87 | var ( |
| 88 | inFlightSem chan struct{} |
| 89 | errCnt = prometheus.NewCounterVec( |
| 90 | prometheus.CounterOpts{ |
| 91 | Name: "promhttp_metric_handler_errors_total", |
| 92 | Help: "Total number of internal errors encountered by the promhttp metric handler.", |
| 93 | }, |
| 94 | []string{"cause"}, |
| 95 | ) |
| 96 | ) |
| 97 | |
| 98 | if opts.MaxRequestsInFlight > 0 { |
| 99 | inFlightSem = make(chan struct{}, opts.MaxRequestsInFlight) |
| 100 | } |
| 101 | if opts.Registry != nil { |
khenaidoo | d948f77 | 2021-08-11 17:49:24 -0400 | [diff] [blame^] | 102 | // Initialize all possibilities that can occur below. |
khenaidoo | ab1f7bd | 2019-11-14 14:00:27 -0500 | [diff] [blame] | 103 | errCnt.WithLabelValues("gathering") |
| 104 | errCnt.WithLabelValues("encoding") |
| 105 | if err := opts.Registry.Register(errCnt); err != nil { |
| 106 | if are, ok := err.(prometheus.AlreadyRegisteredError); ok { |
| 107 | errCnt = are.ExistingCollector.(*prometheus.CounterVec) |
| 108 | } else { |
| 109 | panic(err) |
| 110 | } |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | h := http.HandlerFunc(func(rsp http.ResponseWriter, req *http.Request) { |
| 115 | if inFlightSem != nil { |
| 116 | select { |
| 117 | case inFlightSem <- struct{}{}: // All good, carry on. |
| 118 | defer func() { <-inFlightSem }() |
| 119 | default: |
| 120 | http.Error(rsp, fmt.Sprintf( |
| 121 | "Limit of concurrent requests reached (%d), try again later.", opts.MaxRequestsInFlight, |
| 122 | ), http.StatusServiceUnavailable) |
| 123 | return |
| 124 | } |
| 125 | } |
| 126 | mfs, err := reg.Gather() |
| 127 | if err != nil { |
| 128 | if opts.ErrorLog != nil { |
| 129 | opts.ErrorLog.Println("error gathering metrics:", err) |
| 130 | } |
| 131 | errCnt.WithLabelValues("gathering").Inc() |
| 132 | switch opts.ErrorHandling { |
| 133 | case PanicOnError: |
| 134 | panic(err) |
| 135 | case ContinueOnError: |
| 136 | if len(mfs) == 0 { |
| 137 | // Still report the error if no metrics have been gathered. |
| 138 | httpError(rsp, err) |
| 139 | return |
| 140 | } |
| 141 | case HTTPErrorOnError: |
| 142 | httpError(rsp, err) |
| 143 | return |
| 144 | } |
| 145 | } |
| 146 | |
khenaidoo | d948f77 | 2021-08-11 17:49:24 -0400 | [diff] [blame^] | 147 | var contentType expfmt.Format |
| 148 | if opts.EnableOpenMetrics { |
| 149 | contentType = expfmt.NegotiateIncludingOpenMetrics(req.Header) |
| 150 | } else { |
| 151 | contentType = expfmt.Negotiate(req.Header) |
| 152 | } |
khenaidoo | ab1f7bd | 2019-11-14 14:00:27 -0500 | [diff] [blame] | 153 | header := rsp.Header() |
| 154 | header.Set(contentTypeHeader, string(contentType)) |
| 155 | |
| 156 | w := io.Writer(rsp) |
| 157 | if !opts.DisableCompression && gzipAccepted(req.Header) { |
| 158 | header.Set(contentEncodingHeader, "gzip") |
| 159 | gz := gzipPool.Get().(*gzip.Writer) |
| 160 | defer gzipPool.Put(gz) |
| 161 | |
| 162 | gz.Reset(w) |
| 163 | defer gz.Close() |
| 164 | |
| 165 | w = gz |
| 166 | } |
| 167 | |
| 168 | enc := expfmt.NewEncoder(w, contentType) |
| 169 | |
khenaidoo | d948f77 | 2021-08-11 17:49:24 -0400 | [diff] [blame^] | 170 | // handleError handles the error according to opts.ErrorHandling |
| 171 | // and returns true if we have to abort after the handling. |
| 172 | handleError := func(err error) bool { |
| 173 | if err == nil { |
| 174 | return false |
khenaidoo | ab1f7bd | 2019-11-14 14:00:27 -0500 | [diff] [blame] | 175 | } |
khenaidoo | d948f77 | 2021-08-11 17:49:24 -0400 | [diff] [blame^] | 176 | if opts.ErrorLog != nil { |
| 177 | opts.ErrorLog.Println("error encoding and sending metric family:", err) |
| 178 | } |
| 179 | errCnt.WithLabelValues("encoding").Inc() |
| 180 | switch opts.ErrorHandling { |
| 181 | case PanicOnError: |
| 182 | panic(err) |
| 183 | case HTTPErrorOnError: |
| 184 | // We cannot really send an HTTP error at this |
| 185 | // point because we most likely have written |
| 186 | // something to rsp already. But at least we can |
| 187 | // stop sending. |
| 188 | return true |
| 189 | } |
| 190 | // Do nothing in all other cases, including ContinueOnError. |
| 191 | return false |
khenaidoo | ab1f7bd | 2019-11-14 14:00:27 -0500 | [diff] [blame] | 192 | } |
| 193 | |
khenaidoo | d948f77 | 2021-08-11 17:49:24 -0400 | [diff] [blame^] | 194 | for _, mf := range mfs { |
| 195 | if handleError(enc.Encode(mf)) { |
| 196 | return |
| 197 | } |
| 198 | } |
| 199 | if closer, ok := enc.(expfmt.Closer); ok { |
| 200 | // This in particular takes care of the final "# EOF\n" line for OpenMetrics. |
| 201 | if handleError(closer.Close()) { |
| 202 | return |
| 203 | } |
khenaidoo | ab1f7bd | 2019-11-14 14:00:27 -0500 | [diff] [blame] | 204 | } |
| 205 | }) |
| 206 | |
| 207 | if opts.Timeout <= 0 { |
| 208 | return h |
| 209 | } |
| 210 | return http.TimeoutHandler(h, opts.Timeout, fmt.Sprintf( |
| 211 | "Exceeded configured timeout of %v.\n", |
| 212 | opts.Timeout, |
| 213 | )) |
| 214 | } |
| 215 | |
| 216 | // InstrumentMetricHandler is usually used with an http.Handler returned by the |
| 217 | // HandlerFor function. It instruments the provided http.Handler with two |
| 218 | // metrics: A counter vector "promhttp_metric_handler_requests_total" to count |
| 219 | // scrapes partitioned by HTTP status code, and a gauge |
| 220 | // "promhttp_metric_handler_requests_in_flight" to track the number of |
| 221 | // simultaneous scrapes. This function idempotently registers collectors for |
| 222 | // both metrics with the provided Registerer. It panics if the registration |
| 223 | // fails. The provided metrics are useful to see how many scrapes hit the |
| 224 | // monitored target (which could be from different Prometheus servers or other |
| 225 | // scrapers), and how often they overlap (which would result in more than one |
| 226 | // scrape in flight at the same time). Note that the scrapes-in-flight gauge |
| 227 | // will contain the scrape by which it is exposed, while the scrape counter will |
| 228 | // only get incremented after the scrape is complete (as only then the status |
| 229 | // code is known). For tracking scrape durations, use the |
| 230 | // "scrape_duration_seconds" gauge created by the Prometheus server upon each |
| 231 | // scrape. |
| 232 | func InstrumentMetricHandler(reg prometheus.Registerer, handler http.Handler) http.Handler { |
| 233 | cnt := prometheus.NewCounterVec( |
| 234 | prometheus.CounterOpts{ |
| 235 | Name: "promhttp_metric_handler_requests_total", |
| 236 | Help: "Total number of scrapes by HTTP status code.", |
| 237 | }, |
| 238 | []string{"code"}, |
| 239 | ) |
| 240 | // Initialize the most likely HTTP status codes. |
| 241 | cnt.WithLabelValues("200") |
| 242 | cnt.WithLabelValues("500") |
| 243 | cnt.WithLabelValues("503") |
| 244 | if err := reg.Register(cnt); err != nil { |
| 245 | if are, ok := err.(prometheus.AlreadyRegisteredError); ok { |
| 246 | cnt = are.ExistingCollector.(*prometheus.CounterVec) |
| 247 | } else { |
| 248 | panic(err) |
| 249 | } |
| 250 | } |
| 251 | |
| 252 | gge := prometheus.NewGauge(prometheus.GaugeOpts{ |
| 253 | Name: "promhttp_metric_handler_requests_in_flight", |
| 254 | Help: "Current number of scrapes being served.", |
| 255 | }) |
| 256 | if err := reg.Register(gge); err != nil { |
| 257 | if are, ok := err.(prometheus.AlreadyRegisteredError); ok { |
| 258 | gge = are.ExistingCollector.(prometheus.Gauge) |
| 259 | } else { |
| 260 | panic(err) |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | return InstrumentHandlerCounter(cnt, InstrumentHandlerInFlight(gge, handler)) |
| 265 | } |
| 266 | |
| 267 | // HandlerErrorHandling defines how a Handler serving metrics will handle |
| 268 | // errors. |
| 269 | type HandlerErrorHandling int |
| 270 | |
| 271 | // These constants cause handlers serving metrics to behave as described if |
| 272 | // errors are encountered. |
| 273 | const ( |
| 274 | // Serve an HTTP status code 500 upon the first error |
khenaidoo | d948f77 | 2021-08-11 17:49:24 -0400 | [diff] [blame^] | 275 | // encountered. Report the error message in the body. Note that HTTP |
| 276 | // errors cannot be served anymore once the beginning of a regular |
| 277 | // payload has been sent. Thus, in the (unlikely) case that encoding the |
| 278 | // payload into the negotiated wire format fails, serving the response |
| 279 | // will simply be aborted. Set an ErrorLog in HandlerOpts to detect |
| 280 | // those errors. |
khenaidoo | ab1f7bd | 2019-11-14 14:00:27 -0500 | [diff] [blame] | 281 | HTTPErrorOnError HandlerErrorHandling = iota |
| 282 | // Ignore errors and try to serve as many metrics as possible. However, |
| 283 | // if no metrics can be served, serve an HTTP status code 500 and the |
| 284 | // last error message in the body. Only use this in deliberate "best |
| 285 | // effort" metrics collection scenarios. In this case, it is highly |
| 286 | // recommended to provide other means of detecting errors: By setting an |
| 287 | // ErrorLog in HandlerOpts, the errors are logged. By providing a |
| 288 | // Registry in HandlerOpts, the exposed metrics include an error counter |
| 289 | // "promhttp_metric_handler_errors_total", which can be used for |
| 290 | // alerts. |
| 291 | ContinueOnError |
| 292 | // Panic upon the first error encountered (useful for "crash only" apps). |
| 293 | PanicOnError |
| 294 | ) |
| 295 | |
| 296 | // Logger is the minimal interface HandlerOpts needs for logging. Note that |
| 297 | // log.Logger from the standard library implements this interface, and it is |
| 298 | // easy to implement by custom loggers, if they don't do so already anyway. |
| 299 | type Logger interface { |
| 300 | Println(v ...interface{}) |
| 301 | } |
| 302 | |
| 303 | // HandlerOpts specifies options how to serve metrics via an http.Handler. The |
| 304 | // zero value of HandlerOpts is a reasonable default. |
| 305 | type HandlerOpts struct { |
khenaidoo | d948f77 | 2021-08-11 17:49:24 -0400 | [diff] [blame^] | 306 | // ErrorLog specifies an optional Logger for errors collecting and |
| 307 | // serving metrics. If nil, errors are not logged at all. Note that the |
| 308 | // type of a reported error is often prometheus.MultiError, which |
| 309 | // formats into a multi-line error string. If you want to avoid the |
| 310 | // latter, create a Logger implementation that detects a |
| 311 | // prometheus.MultiError and formats the contained errors into one line. |
khenaidoo | ab1f7bd | 2019-11-14 14:00:27 -0500 | [diff] [blame] | 312 | ErrorLog Logger |
| 313 | // ErrorHandling defines how errors are handled. Note that errors are |
| 314 | // logged regardless of the configured ErrorHandling provided ErrorLog |
| 315 | // is not nil. |
| 316 | ErrorHandling HandlerErrorHandling |
| 317 | // If Registry is not nil, it is used to register a metric |
| 318 | // "promhttp_metric_handler_errors_total", partitioned by "cause". A |
| 319 | // failed registration causes a panic. Note that this error counter is |
| 320 | // different from the instrumentation you get from the various |
| 321 | // InstrumentHandler... helpers. It counts errors that don't necessarily |
| 322 | // result in a non-2xx HTTP status code. There are two typical cases: |
| 323 | // (1) Encoding errors that only happen after streaming of the HTTP body |
| 324 | // has already started (and the status code 200 has been sent). This |
| 325 | // should only happen with custom collectors. (2) Collection errors with |
| 326 | // no effect on the HTTP status code because ErrorHandling is set to |
| 327 | // ContinueOnError. |
| 328 | Registry prometheus.Registerer |
| 329 | // If DisableCompression is true, the handler will never compress the |
| 330 | // response, even if requested by the client. |
| 331 | DisableCompression bool |
| 332 | // The number of concurrent HTTP requests is limited to |
| 333 | // MaxRequestsInFlight. Additional requests are responded to with 503 |
| 334 | // Service Unavailable and a suitable message in the body. If |
| 335 | // MaxRequestsInFlight is 0 or negative, no limit is applied. |
| 336 | MaxRequestsInFlight int |
| 337 | // If handling a request takes longer than Timeout, it is responded to |
| 338 | // with 503 ServiceUnavailable and a suitable Message. No timeout is |
| 339 | // applied if Timeout is 0 or negative. Note that with the current |
| 340 | // implementation, reaching the timeout simply ends the HTTP requests as |
| 341 | // described above (and even that only if sending of the body hasn't |
| 342 | // started yet), while the bulk work of gathering all the metrics keeps |
| 343 | // running in the background (with the eventual result to be thrown |
| 344 | // away). Until the implementation is improved, it is recommended to |
| 345 | // implement a separate timeout in potentially slow Collectors. |
| 346 | Timeout time.Duration |
khenaidoo | d948f77 | 2021-08-11 17:49:24 -0400 | [diff] [blame^] | 347 | // If true, the experimental OpenMetrics encoding is added to the |
| 348 | // possible options during content negotiation. Note that Prometheus |
| 349 | // 2.5.0+ will negotiate OpenMetrics as first priority. OpenMetrics is |
| 350 | // the only way to transmit exemplars. However, the move to OpenMetrics |
| 351 | // is not completely transparent. Most notably, the values of "quantile" |
| 352 | // labels of Summaries and "le" labels of Histograms are formatted with |
| 353 | // a trailing ".0" if they would otherwise look like integer numbers |
| 354 | // (which changes the identity of the resulting series on the Prometheus |
| 355 | // server). |
| 356 | EnableOpenMetrics bool |
khenaidoo | ab1f7bd | 2019-11-14 14:00:27 -0500 | [diff] [blame] | 357 | } |
| 358 | |
| 359 | // gzipAccepted returns whether the client will accept gzip-encoded content. |
| 360 | func gzipAccepted(header http.Header) bool { |
| 361 | a := header.Get(acceptEncodingHeader) |
| 362 | parts := strings.Split(a, ",") |
| 363 | for _, part := range parts { |
| 364 | part = strings.TrimSpace(part) |
| 365 | if part == "gzip" || strings.HasPrefix(part, "gzip;") { |
| 366 | return true |
| 367 | } |
| 368 | } |
| 369 | return false |
| 370 | } |
| 371 | |
| 372 | // httpError removes any content-encoding header and then calls http.Error with |
khenaidoo | d948f77 | 2021-08-11 17:49:24 -0400 | [diff] [blame^] | 373 | // the provided error and http.StatusInternalServerError. Error contents is |
| 374 | // supposed to be uncompressed plain text. Same as with a plain http.Error, this |
| 375 | // must not be called if the header or any payload has already been sent. |
khenaidoo | ab1f7bd | 2019-11-14 14:00:27 -0500 | [diff] [blame] | 376 | func httpError(rsp http.ResponseWriter, err error) { |
| 377 | rsp.Header().Del(contentEncodingHeader) |
| 378 | http.Error( |
| 379 | rsp, |
| 380 | "An error has occurred while serving metrics:\n\n"+err.Error(), |
| 381 | http.StatusInternalServerError, |
| 382 | ) |
| 383 | } |