1 // Copyright 2014 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
6 // http://www.apache.org/licenses/LICENSE-2.0
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.
25 dto "github.com/prometheus/client_model/go"
27 "github.com/golang/protobuf/proto"
28 "github.com/prometheus/common/model"
31 // A stateFn is a function that represents a state in a state machine. By
32 // executing it, the state is progressed to the next state. The stateFn returns
33 // another stateFn, which represents the new state. The end state is represented
35 type stateFn func() stateFn
37 // ParseError signals errors while parsing the simple and flat text-based
39 type ParseError struct {
44 // Error implements the error interface.
45 func (e ParseError) Error() string {
46 return fmt.Sprintf("text format parsing error in line %d: %s", e.Line, e.Msg)
49 // TextParser is used to parse the simple and flat text-based exchange format. Its
50 // nil value is ready to use.
51 type TextParser struct {
52 metricFamiliesByName map[string]*dto.MetricFamily
53 buf *bufio.Reader // Where the parsed input is read through.
54 err error // Most recent error.
55 lineCount int // Tracks the line count for error messages.
56 currentByte byte // The most recent byte read.
57 currentToken bytes.Buffer // Re-used each time a token has to be gathered from multiple bytes.
58 currentMF *dto.MetricFamily
59 currentMetric *dto.Metric
60 currentLabelPair *dto.LabelPair
62 // The remaining member variables are only used for summaries/histograms.
63 currentLabels map[string]string // All labels including '__name__' but excluding 'quantile'/'le'
65 summaries map[uint64]*dto.Metric // Key is created with LabelsToSignature.
66 currentQuantile float64
67 // Histogram specific.
68 histograms map[uint64]*dto.Metric // Key is created with LabelsToSignature.
70 // These tell us if the currently processed line ends on '_count' or
71 // '_sum' respectively and belong to a summary/histogram, representing the sample
72 // count and sum of that summary/histogram.
73 currentIsSummaryCount, currentIsSummarySum bool
74 currentIsHistogramCount, currentIsHistogramSum bool
77 // TextToMetricFamilies reads 'in' as the simple and flat text-based exchange
78 // format and creates MetricFamily proto messages. It returns the MetricFamily
79 // proto messages in a map where the metric names are the keys, along with any
82 // If the input contains duplicate metrics (i.e. lines with the same metric name
83 // and exactly the same label set), the resulting MetricFamily will contain
84 // duplicate Metric proto messages. Similar is true for duplicate label
85 // names. Checks for duplicates have to be performed separately, if required.
86 // Also note that neither the metrics within each MetricFamily are sorted nor
87 // the label pairs within each Metric. Sorting is not required for the most
88 // frequent use of this method, which is sample ingestion in the Prometheus
89 // server. However, for presentation purposes, you might want to sort the
90 // metrics, and in some cases, you must sort the labels, e.g. for consumption by
91 // the metric family injection hook of the Prometheus registry.
93 // Summaries and histograms are rather special beasts. You would probably not
94 // use them in the simple text format anyway. This method can deal with
95 // summaries and histograms if they are presented in exactly the way the
96 // text.Create function creates them.
98 // This method must not be called concurrently. If you want to parse different
99 // input concurrently, instantiate a separate Parser for each goroutine.
100 func (p *TextParser) TextToMetricFamilies(in io.Reader) (map[string]*dto.MetricFamily, error) {
102 for nextState := p.startOfLine; nextState != nil; nextState = nextState() {
103 // Magic happens here...
105 // Get rid of empty metric families.
106 for k, mf := range p.metricFamiliesByName {
107 if len(mf.GetMetric()) == 0 {
108 delete(p.metricFamiliesByName, k)
111 // If p.err is io.EOF now, we have run into a premature end of the input
112 // stream. Turn this error into something nicer and more
113 // meaningful. (io.EOF is often used as a signal for the legitimate end
114 // of an input stream.)
116 p.parseError("unexpected end of input stream")
118 return p.metricFamiliesByName, p.err
121 func (p *TextParser) reset(in io.Reader) {
122 p.metricFamiliesByName = map[string]*dto.MetricFamily{}
124 p.buf = bufio.NewReader(in)
130 if p.summaries == nil || len(p.summaries) > 0 {
131 p.summaries = map[uint64]*dto.Metric{}
133 if p.histograms == nil || len(p.histograms) > 0 {
134 p.histograms = map[uint64]*dto.Metric{}
136 p.currentQuantile = math.NaN()
137 p.currentBucket = math.NaN()
140 // startOfLine represents the state where the next byte read from p.buf is the
141 // start of a line (or whitespace leading up to it).
142 func (p *TextParser) startOfLine() stateFn {
144 if p.skipBlankTab(); p.err != nil {
145 // End of input reached. This is the only case where
146 // that is not an error but a signal that we are done.
150 switch p.currentByte {
152 return p.startComment
154 return p.startOfLine // Empty line, start the next one.
156 return p.readingMetricName
159 // startComment represents the state where the next byte read from p.buf is the
160 // start of a comment (or whitespace leading up to it).
161 func (p *TextParser) startComment() stateFn {
162 if p.skipBlankTab(); p.err != nil {
163 return nil // Unexpected end of input.
165 if p.currentByte == '\n' {
168 if p.readTokenUntilWhitespace(); p.err != nil {
169 return nil // Unexpected end of input.
171 // If we have hit the end of line already, there is nothing left
172 // to do. This is not considered a syntax error.
173 if p.currentByte == '\n' {
176 keyword := p.currentToken.String()
177 if keyword != "HELP" && keyword != "TYPE" {
178 // Generic comment, ignore by fast forwarding to end of line.
179 for p.currentByte != '\n' {
180 if p.currentByte, p.err = p.buf.ReadByte(); p.err != nil {
181 return nil // Unexpected end of input.
186 // There is something. Next has to be a metric name.
187 if p.skipBlankTab(); p.err != nil {
188 return nil // Unexpected end of input.
190 if p.readTokenAsMetricName(); p.err != nil {
191 return nil // Unexpected end of input.
193 if p.currentByte == '\n' {
194 // At the end of the line already.
195 // Again, this is not considered a syntax error.
198 if !isBlankOrTab(p.currentByte) {
199 p.parseError("invalid metric name in comment")
202 p.setOrCreateCurrentMF()
203 if p.skipBlankTab(); p.err != nil {
204 return nil // Unexpected end of input.
206 if p.currentByte == '\n' {
207 // At the end of the line already.
208 // Again, this is not considered a syntax error.
217 panic(fmt.Sprintf("code error: unexpected keyword %q", keyword))
220 // readingMetricName represents the state where the last byte read (now in
221 // p.currentByte) is the first byte of a metric name.
222 func (p *TextParser) readingMetricName() stateFn {
223 if p.readTokenAsMetricName(); p.err != nil {
226 if p.currentToken.Len() == 0 {
227 p.parseError("invalid metric name")
230 p.setOrCreateCurrentMF()
231 // Now is the time to fix the type if it hasn't happened yet.
232 if p.currentMF.Type == nil {
233 p.currentMF.Type = dto.MetricType_UNTYPED.Enum()
235 p.currentMetric = &dto.Metric{}
236 // Do not append the newly created currentMetric to
237 // currentMF.Metric right now. First wait if this is a summary,
238 // and the metric exists already, which we can only know after
239 // having read all the labels.
240 if p.skipBlankTabIfCurrentBlankTab(); p.err != nil {
241 return nil // Unexpected end of input.
243 return p.readingLabels
246 // readingLabels represents the state where the last byte read (now in
247 // p.currentByte) is either the first byte of the label set (i.e. a '{'), or the
248 // first byte of the value (otherwise).
249 func (p *TextParser) readingLabels() stateFn {
250 // Summaries/histograms are special. We have to reset the
251 // currentLabels map, currentQuantile and currentBucket before starting to
253 if p.currentMF.GetType() == dto.MetricType_SUMMARY || p.currentMF.GetType() == dto.MetricType_HISTOGRAM {
254 p.currentLabels = map[string]string{}
255 p.currentLabels[string(model.MetricNameLabel)] = p.currentMF.GetName()
256 p.currentQuantile = math.NaN()
257 p.currentBucket = math.NaN()
259 if p.currentByte != '{' {
260 return p.readingValue
262 return p.startLabelName
265 // startLabelName represents the state where the next byte read from p.buf is
266 // the start of a label name (or whitespace leading up to it).
267 func (p *TextParser) startLabelName() stateFn {
268 if p.skipBlankTab(); p.err != nil {
269 return nil // Unexpected end of input.
271 if p.currentByte == '}' {
272 if p.skipBlankTab(); p.err != nil {
273 return nil // Unexpected end of input.
275 return p.readingValue
277 if p.readTokenAsLabelName(); p.err != nil {
278 return nil // Unexpected end of input.
280 if p.currentToken.Len() == 0 {
281 p.parseError(fmt.Sprintf("invalid label name for metric %q", p.currentMF.GetName()))
284 p.currentLabelPair = &dto.LabelPair{Name: proto.String(p.currentToken.String())}
285 if p.currentLabelPair.GetName() == string(model.MetricNameLabel) {
286 p.parseError(fmt.Sprintf("label name %q is reserved", model.MetricNameLabel))
289 // Special summary/histogram treatment. Don't add 'quantile' and 'le'
290 // labels to 'real' labels.
291 if !(p.currentMF.GetType() == dto.MetricType_SUMMARY && p.currentLabelPair.GetName() == model.QuantileLabel) &&
292 !(p.currentMF.GetType() == dto.MetricType_HISTOGRAM && p.currentLabelPair.GetName() == model.BucketLabel) {
293 p.currentMetric.Label = append(p.currentMetric.Label, p.currentLabelPair)
295 if p.skipBlankTabIfCurrentBlankTab(); p.err != nil {
296 return nil // Unexpected end of input.
298 if p.currentByte != '=' {
299 p.parseError(fmt.Sprintf("expected '=' after label name, found %q", p.currentByte))
302 return p.startLabelValue
305 // startLabelValue represents the state where the next byte read from p.buf is
306 // the start of a (quoted) label value (or whitespace leading up to it).
307 func (p *TextParser) startLabelValue() stateFn {
308 if p.skipBlankTab(); p.err != nil {
309 return nil // Unexpected end of input.
311 if p.currentByte != '"' {
312 p.parseError(fmt.Sprintf("expected '\"' at start of label value, found %q", p.currentByte))
315 if p.readTokenAsLabelValue(); p.err != nil {
318 p.currentLabelPair.Value = proto.String(p.currentToken.String())
319 // Special treatment of summaries:
320 // - Quantile labels are special, will result in dto.Quantile later.
321 // - Other labels have to be added to currentLabels for signature calculation.
322 if p.currentMF.GetType() == dto.MetricType_SUMMARY {
323 if p.currentLabelPair.GetName() == model.QuantileLabel {
324 if p.currentQuantile, p.err = strconv.ParseFloat(p.currentLabelPair.GetValue(), 64); p.err != nil {
325 // Create a more helpful error message.
326 p.parseError(fmt.Sprintf("expected float as value for 'quantile' label, got %q", p.currentLabelPair.GetValue()))
330 p.currentLabels[p.currentLabelPair.GetName()] = p.currentLabelPair.GetValue()
333 // Similar special treatment of histograms.
334 if p.currentMF.GetType() == dto.MetricType_HISTOGRAM {
335 if p.currentLabelPair.GetName() == model.BucketLabel {
336 if p.currentBucket, p.err = strconv.ParseFloat(p.currentLabelPair.GetValue(), 64); p.err != nil {
337 // Create a more helpful error message.
338 p.parseError(fmt.Sprintf("expected float as value for 'le' label, got %q", p.currentLabelPair.GetValue()))
342 p.currentLabels[p.currentLabelPair.GetName()] = p.currentLabelPair.GetValue()
345 if p.skipBlankTab(); p.err != nil {
346 return nil // Unexpected end of input.
348 switch p.currentByte {
350 return p.startLabelName
353 if p.skipBlankTab(); p.err != nil {
354 return nil // Unexpected end of input.
356 return p.readingValue
358 p.parseError(fmt.Sprintf("unexpected end of label value %q", p.currentLabelPair.Value))
363 // readingValue represents the state where the last byte read (now in
364 // p.currentByte) is the first byte of the sample value (i.e. a float).
365 func (p *TextParser) readingValue() stateFn {
366 // When we are here, we have read all the labels, so for the
367 // special case of a summary/histogram, we can finally find out
368 // if the metric already exists.
369 if p.currentMF.GetType() == dto.MetricType_SUMMARY {
370 signature := model.LabelsToSignature(p.currentLabels)
371 if summary := p.summaries[signature]; summary != nil {
372 p.currentMetric = summary
374 p.summaries[signature] = p.currentMetric
375 p.currentMF.Metric = append(p.currentMF.Metric, p.currentMetric)
377 } else if p.currentMF.GetType() == dto.MetricType_HISTOGRAM {
378 signature := model.LabelsToSignature(p.currentLabels)
379 if histogram := p.histograms[signature]; histogram != nil {
380 p.currentMetric = histogram
382 p.histograms[signature] = p.currentMetric
383 p.currentMF.Metric = append(p.currentMF.Metric, p.currentMetric)
386 p.currentMF.Metric = append(p.currentMF.Metric, p.currentMetric)
388 if p.readTokenUntilWhitespace(); p.err != nil {
389 return nil // Unexpected end of input.
391 value, err := strconv.ParseFloat(p.currentToken.String(), 64)
393 // Create a more helpful error message.
394 p.parseError(fmt.Sprintf("expected float as value, got %q", p.currentToken.String()))
397 switch p.currentMF.GetType() {
398 case dto.MetricType_COUNTER:
399 p.currentMetric.Counter = &dto.Counter{Value: proto.Float64(value)}
400 case dto.MetricType_GAUGE:
401 p.currentMetric.Gauge = &dto.Gauge{Value: proto.Float64(value)}
402 case dto.MetricType_UNTYPED:
403 p.currentMetric.Untyped = &dto.Untyped{Value: proto.Float64(value)}
404 case dto.MetricType_SUMMARY:
406 if p.currentMetric.Summary == nil {
407 p.currentMetric.Summary = &dto.Summary{}
410 case p.currentIsSummaryCount:
411 p.currentMetric.Summary.SampleCount = proto.Uint64(uint64(value))
412 case p.currentIsSummarySum:
413 p.currentMetric.Summary.SampleSum = proto.Float64(value)
414 case !math.IsNaN(p.currentQuantile):
415 p.currentMetric.Summary.Quantile = append(
416 p.currentMetric.Summary.Quantile,
418 Quantile: proto.Float64(p.currentQuantile),
419 Value: proto.Float64(value),
423 case dto.MetricType_HISTOGRAM:
425 if p.currentMetric.Histogram == nil {
426 p.currentMetric.Histogram = &dto.Histogram{}
429 case p.currentIsHistogramCount:
430 p.currentMetric.Histogram.SampleCount = proto.Uint64(uint64(value))
431 case p.currentIsHistogramSum:
432 p.currentMetric.Histogram.SampleSum = proto.Float64(value)
433 case !math.IsNaN(p.currentBucket):
434 p.currentMetric.Histogram.Bucket = append(
435 p.currentMetric.Histogram.Bucket,
437 UpperBound: proto.Float64(p.currentBucket),
438 CumulativeCount: proto.Uint64(uint64(value)),
443 p.err = fmt.Errorf("unexpected type for metric name %q", p.currentMF.GetName())
445 if p.currentByte == '\n' {
448 return p.startTimestamp
451 // startTimestamp represents the state where the next byte read from p.buf is
452 // the start of the timestamp (or whitespace leading up to it).
453 func (p *TextParser) startTimestamp() stateFn {
454 if p.skipBlankTab(); p.err != nil {
455 return nil // Unexpected end of input.
457 if p.readTokenUntilWhitespace(); p.err != nil {
458 return nil // Unexpected end of input.
460 timestamp, err := strconv.ParseInt(p.currentToken.String(), 10, 64)
462 // Create a more helpful error message.
463 p.parseError(fmt.Sprintf("expected integer as timestamp, got %q", p.currentToken.String()))
466 p.currentMetric.TimestampMs = proto.Int64(timestamp)
467 if p.readTokenUntilNewline(false); p.err != nil {
468 return nil // Unexpected end of input.
470 if p.currentToken.Len() > 0 {
471 p.parseError(fmt.Sprintf("spurious string after timestamp: %q", p.currentToken.String()))
477 // readingHelp represents the state where the last byte read (now in
478 // p.currentByte) is the first byte of the docstring after 'HELP'.
479 func (p *TextParser) readingHelp() stateFn {
480 if p.currentMF.Help != nil {
481 p.parseError(fmt.Sprintf("second HELP line for metric name %q", p.currentMF.GetName()))
484 // Rest of line is the docstring.
485 if p.readTokenUntilNewline(true); p.err != nil {
486 return nil // Unexpected end of input.
488 p.currentMF.Help = proto.String(p.currentToken.String())
492 // readingType represents the state where the last byte read (now in
493 // p.currentByte) is the first byte of the type hint after 'HELP'.
494 func (p *TextParser) readingType() stateFn {
495 if p.currentMF.Type != nil {
496 p.parseError(fmt.Sprintf("second TYPE line for metric name %q, or TYPE reported after samples", p.currentMF.GetName()))
499 // Rest of line is the type.
500 if p.readTokenUntilNewline(false); p.err != nil {
501 return nil // Unexpected end of input.
503 metricType, ok := dto.MetricType_value[strings.ToUpper(p.currentToken.String())]
505 p.parseError(fmt.Sprintf("unknown metric type %q", p.currentToken.String()))
508 p.currentMF.Type = dto.MetricType(metricType).Enum()
512 // parseError sets p.err to a ParseError at the current line with the given
514 func (p *TextParser) parseError(msg string) {
521 // skipBlankTab reads (and discards) bytes from p.buf until it encounters a byte
522 // that is neither ' ' nor '\t'. That byte is left in p.currentByte.
523 func (p *TextParser) skipBlankTab() {
525 if p.currentByte, p.err = p.buf.ReadByte(); p.err != nil || !isBlankOrTab(p.currentByte) {
531 // skipBlankTabIfCurrentBlankTab works exactly as skipBlankTab but doesn't do
532 // anything if p.currentByte is neither ' ' nor '\t'.
533 func (p *TextParser) skipBlankTabIfCurrentBlankTab() {
534 if isBlankOrTab(p.currentByte) {
539 // readTokenUntilWhitespace copies bytes from p.buf into p.currentToken. The
540 // first byte considered is the byte already read (now in p.currentByte). The
541 // first whitespace byte encountered is still copied into p.currentByte, but not
542 // into p.currentToken.
543 func (p *TextParser) readTokenUntilWhitespace() {
544 p.currentToken.Reset()
545 for p.err == nil && !isBlankOrTab(p.currentByte) && p.currentByte != '\n' {
546 p.currentToken.WriteByte(p.currentByte)
547 p.currentByte, p.err = p.buf.ReadByte()
551 // readTokenUntilNewline copies bytes from p.buf into p.currentToken. The first
552 // byte considered is the byte already read (now in p.currentByte). The first
553 // newline byte encountered is still copied into p.currentByte, but not into
554 // p.currentToken. If recognizeEscapeSequence is true, two escape sequences are
555 // recognized: '\\' tranlates into '\', and '\n' into a line-feed character. All
556 // other escape sequences are invalid and cause an error.
557 func (p *TextParser) readTokenUntilNewline(recognizeEscapeSequence bool) {
558 p.currentToken.Reset()
561 if recognizeEscapeSequence && escaped {
562 switch p.currentByte {
564 p.currentToken.WriteByte(p.currentByte)
566 p.currentToken.WriteByte('\n')
568 p.parseError(fmt.Sprintf("invalid escape sequence '\\%c'", p.currentByte))
573 switch p.currentByte {
579 p.currentToken.WriteByte(p.currentByte)
582 p.currentByte, p.err = p.buf.ReadByte()
586 // readTokenAsMetricName copies a metric name from p.buf into p.currentToken.
587 // The first byte considered is the byte already read (now in p.currentByte).
588 // The first byte not part of a metric name is still copied into p.currentByte,
589 // but not into p.currentToken.
590 func (p *TextParser) readTokenAsMetricName() {
591 p.currentToken.Reset()
592 if !isValidMetricNameStart(p.currentByte) {
596 p.currentToken.WriteByte(p.currentByte)
597 p.currentByte, p.err = p.buf.ReadByte()
598 if p.err != nil || !isValidMetricNameContinuation(p.currentByte) {
604 // readTokenAsLabelName copies a label name from p.buf into p.currentToken.
605 // The first byte considered is the byte already read (now in p.currentByte).
606 // The first byte not part of a label name is still copied into p.currentByte,
607 // but not into p.currentToken.
608 func (p *TextParser) readTokenAsLabelName() {
609 p.currentToken.Reset()
610 if !isValidLabelNameStart(p.currentByte) {
614 p.currentToken.WriteByte(p.currentByte)
615 p.currentByte, p.err = p.buf.ReadByte()
616 if p.err != nil || !isValidLabelNameContinuation(p.currentByte) {
622 // readTokenAsLabelValue copies a label value from p.buf into p.currentToken.
623 // In contrast to the other 'readTokenAs...' functions, which start with the
624 // last read byte in p.currentByte, this method ignores p.currentByte and starts
625 // with reading a new byte from p.buf. The first byte not part of a label value
626 // is still copied into p.currentByte, but not into p.currentToken.
627 func (p *TextParser) readTokenAsLabelValue() {
628 p.currentToken.Reset()
631 if p.currentByte, p.err = p.buf.ReadByte(); p.err != nil {
635 switch p.currentByte {
637 p.currentToken.WriteByte(p.currentByte)
639 p.currentToken.WriteByte('\n')
641 p.parseError(fmt.Sprintf("invalid escape sequence '\\%c'", p.currentByte))
647 switch p.currentByte {
651 p.parseError(fmt.Sprintf("label value %q contains unescaped new-line", p.currentToken.String()))
656 p.currentToken.WriteByte(p.currentByte)
661 func (p *TextParser) setOrCreateCurrentMF() {
662 p.currentIsSummaryCount = false
663 p.currentIsSummarySum = false
664 p.currentIsHistogramCount = false
665 p.currentIsHistogramSum = false
666 name := p.currentToken.String()
667 if p.currentMF = p.metricFamiliesByName[name]; p.currentMF != nil {
670 // Try out if this is a _sum or _count for a summary/histogram.
671 summaryName := summaryMetricName(name)
672 if p.currentMF = p.metricFamiliesByName[summaryName]; p.currentMF != nil {
673 if p.currentMF.GetType() == dto.MetricType_SUMMARY {
675 p.currentIsSummaryCount = true
678 p.currentIsSummarySum = true
683 histogramName := histogramMetricName(name)
684 if p.currentMF = p.metricFamiliesByName[histogramName]; p.currentMF != nil {
685 if p.currentMF.GetType() == dto.MetricType_HISTOGRAM {
687 p.currentIsHistogramCount = true
690 p.currentIsHistogramSum = true
695 p.currentMF = &dto.MetricFamily{Name: proto.String(name)}
696 p.metricFamiliesByName[name] = p.currentMF
699 func isValidLabelNameStart(b byte) bool {
700 return (b >= 'a' && b <= 'z') || (b >= 'A' && b <= 'Z') || b == '_'
703 func isValidLabelNameContinuation(b byte) bool {
704 return isValidLabelNameStart(b) || (b >= '0' && b <= '9')
707 func isValidMetricNameStart(b byte) bool {
708 return isValidLabelNameStart(b) || b == ':'
711 func isValidMetricNameContinuation(b byte) bool {
712 return isValidLabelNameContinuation(b) || b == ':'
715 func isBlankOrTab(b byte) bool {
716 return b == ' ' || b == '\t'
719 func isCount(name string) bool {
720 return len(name) > 6 && name[len(name)-6:] == "_count"
723 func isSum(name string) bool {
724 return len(name) > 4 && name[len(name)-4:] == "_sum"
727 func isBucket(name string) bool {
728 return len(name) > 7 && name[len(name)-7:] == "_bucket"
731 func summaryMetricName(name string) string {
734 return name[:len(name)-6]
736 return name[:len(name)-4]
742 func histogramMetricName(name string) string {
745 return name[:len(name)-6]
747 return name[:len(name)-4]
749 return name[:len(name)-7]