grafana/public/app/plugins/datasource/prometheus/result_transformer.ts

import {
  ArrayDataFrame,
  ArrayVector,
  DataFrame,
  DataLink,
  DataTopic,
  Field,
  FieldType,
  formatLabels,
  getDisplayProcessor,
  Labels,
  MutableField,
  ScopedVars,
  TIME_SERIES_TIME_FIELD_NAME,
  TIME_SERIES_VALUE_FIELD_NAME,
} from '@grafana/data';
import { FetchResponse, getDataSourceSrv, getTemplateSrv } from '@grafana/runtime';
import { descending, deviation } from 'd3';
import {
  ExemplarTraceIdDestination,
  isExemplarData,
  isMatrixData,
  MatrixOrVectorResult,
  PromDataSuccessResponse,
  PromMetric,
  PromQuery,
  PromQueryRequest,
  PromValue,
  TransformOptions,
} from './types';

const POSITIVE_INFINITY_SAMPLE_VALUE = '+Inf';
const NEGATIVE_INFINITY_SAMPLE_VALUE = '-Inf';

interface TimeAndValue {
  [TIME_SERIES_TIME_FIELD_NAME]: number;
  [TIME_SERIES_VALUE_FIELD_NAME]: number;
}

export function transform(
  response: FetchResponse<PromDataSuccessResponse>,
  transformOptions: {
    query: PromQueryRequest;
    exemplarTraceIdDestinations?: ExemplarTraceIdDestination[];
    target: PromQuery;
    responseListLength: number;
    scopedVars?: ScopedVars;
  }
) {
  // Create options object from transformOptions
  const options: TransformOptions = {
    format: transformOptions.target.format,
    step: transformOptions.query.step,
    legendFormat: transformOptions.target.legendFormat,
    start: transformOptions.query.start,
    end: transformOptions.query.end,
    query: transformOptions.query.expr,
    responseListLength: transformOptions.responseListLength,
    scopedVars: transformOptions.scopedVars,
    refId: transformOptions.target.refId,
    valueWithRefId: transformOptions.target.valueWithRefId,
    meta: {
      // Fix for showing of Prometheus results in Explore table
      preferredVisualisationType: transformOptions.query.instant ? 'table' : 'graph',
    },
  };
  const prometheusResult = response.data.data;

  if (isExemplarData(prometheusResult)) {
    const events: TimeAndValue[] = [];
    prometheusResult.forEach((exemplarData) => {
      const data = exemplarData.exemplars.map((exemplar) => {
        return {
          [TIME_SERIES_TIME_FIELD_NAME]: exemplar.timestamp * 1000,
          [TIME_SERIES_VALUE_FIELD_NAME]: exemplar.value,
          ...exemplar.labels,
          ...exemplarData.seriesLabels,
        };
      });
      events.push(...data);
    });

    // Grouping exemplars by step
    const sampledExemplars = sampleExemplars(events, options);

    const dataFrame = new ArrayDataFrame(sampledExemplars);
    dataFrame.meta = { dataTopic: DataTopic.Annotations };

    // Add data links if configured
    if (transformOptions.exemplarTraceIdDestinations?.length) {
      for (const exemplarTraceIdDestination of transformOptions.exemplarTraceIdDestinations) {
        const traceIDField = dataFrame.fields.find((field) => field.name === exemplarTraceIdDestination!.name);
        if (traceIDField) {
          const links = getDataLinks(exemplarTraceIdDestination);
          traceIDField.config.links = traceIDField.config.links?.length
            ? [...traceIDField.config.links, ...links]
            : links;
        }
      }
    }
    return [dataFrame];
  }

  if (!prometheusResult?.result) {
    return [];
  }

  // Return early if result type is scalar
  if (prometheusResult.resultType === 'scalar') {
    return [
      {
        meta: options.meta,
        refId: options.refId,
        length: 1,
        fields: [getTimeField([prometheusResult.result]), getValueField({ data: [prometheusResult.result] })],
      },
    ];
  }

  // Return early again if the format is table, this needs special transformation.
  if (options.format === 'table') {
    const tableData = transformMetricDataToTable(prometheusResult.result, options);
    return [tableData];
  }

  // Process matrix and vector results to DataFrame
  const dataFrame: DataFrame[] = [];
  prometheusResult.result.forEach((data: MatrixOrVectorResult) => dataFrame.push(transformToDataFrame(data, options)));

  // When format is heatmap use the already created data frames and transform it more
  if (options.format === 'heatmap') {
    dataFrame.sort(sortSeriesByLabel);
    const seriesList = transformToHistogramOverTime(dataFrame);
    return seriesList;
  }

  // Return matrix or vector result as DataFrame[]
  return dataFrame;
}

function getDataLinks(options: ExemplarTraceIdDestination): DataLink[] {
  const dataLinks: DataLink[] = [];

  if (options.datasourceUid) {
    const dataSourceSrv = getDataSourceSrv();
    const dsSettings = dataSourceSrv.getInstanceSettings(options.datasourceUid);

    dataLinks.push({
      title: `Query with ${dsSettings?.name}`,
      url: '',
      internal: {
        query: { query: '${__value.raw}', queryType: 'traceId' },
        datasourceUid: options.datasourceUid,
        datasourceName: dsSettings?.name ?? 'Data source not found',
      },
    });
  }

  if (options.url) {
    dataLinks.push({
      title: `Go to ${options.url}`,
      url: options.url,
      targetBlank: true,
    });
  }
  return dataLinks;
}

/**
 * Reduce the density of the exemplars by making sure that the highest value exemplar is included
 * and then only the ones that are 2 times the standard deviation of the all the values.
 * This makes sure not to show too many dots near each other.
 */
function sampleExemplars(events: TimeAndValue[], options: TransformOptions) {
  const step = options.step || 15;
  const bucketedExemplars: { [ts: string]: TimeAndValue[] } = {};
  const values: number[] = [];
  for (const exemplar of events) {
    // Align exemplar timestamp to nearest step second
    const alignedTs = String(Math.floor(exemplar[TIME_SERIES_TIME_FIELD_NAME] / 1000 / step) * step * 1000);
    if (!bucketedExemplars[alignedTs]) {
      // New bucket found
      bucketedExemplars[alignedTs] = [];
    }
    bucketedExemplars[alignedTs].push(exemplar);
    values.push(exemplar[TIME_SERIES_VALUE_FIELD_NAME]);
  }

  // Getting exemplars from each bucket
  const standardDeviation = deviation(values);
  const sampledBuckets = Object.keys(bucketedExemplars).sort();
  const sampledExemplars = [];
  for (const ts of sampledBuckets) {
    const exemplarsInBucket = bucketedExemplars[ts];
    if (exemplarsInBucket.length === 1) {
      sampledExemplars.push(exemplarsInBucket[0]);
    } else {
      // Choose which values to sample
      const bucketValues = exemplarsInBucket.map((ex) => ex[TIME_SERIES_VALUE_FIELD_NAME]).sort(descending);
      const sampledBucketValues = bucketValues.reduce((acc: number[], curr) => {
        if (acc.length === 0) {
          // First value is max and is always added
          acc.push(curr);
        } else {
          // Then take values only when at least 2 standard deviation distance to previously taken value
          const prev = acc[acc.length - 1];
          if (standardDeviation && prev - curr >= 2 * standardDeviation) {
            acc.push(curr);
          }
        }
        return acc;
      }, []);
      // Find the exemplars for the sampled values
      sampledExemplars.push(
        ...sampledBucketValues.map(
          (value) => exemplarsInBucket.find((ex) => ex[TIME_SERIES_VALUE_FIELD_NAME] === value)!
        )
      );
    }
  }
  return sampledExemplars;
}

/**
 * Transforms matrix and vector result from Prometheus result to DataFrame
 */
function transformToDataFrame(data: MatrixOrVectorResult, options: TransformOptions): DataFrame {
  const { name, labels } = createLabelInfo(data.metric, options);

  const fields: Field[] = [];

  if (isMatrixData(data)) {
    const stepMs = options.step ? options.step * 1000 : NaN;
    let baseTimestamp = options.start * 1000;
    const dps: PromValue[] = [];

    for (const value of data.values) {
      let dpValue: number | null = parseSampleValue(value[1]);

      if (isNaN(dpValue)) {
        dpValue = null;
      }

      const timestamp = value[0] * 1000;
      for (let t = baseTimestamp; t < timestamp; t += stepMs) {
        dps.push([t, null]);
      }
      baseTimestamp = timestamp + stepMs;
      dps.push([timestamp, dpValue]);
    }

    const endTimestamp = options.end * 1000;
    for (let t = baseTimestamp; t <= endTimestamp; t += stepMs) {
      dps.push([t, null]);
    }
    fields.push(getTimeField(dps, true));
    fields.push(getValueField({ data: dps, parseValue: false, labels, displayNameFromDS: name }));
  } else {
    fields.push(getTimeField([data.value]));
    fields.push(getValueField({ data: [data.value], labels, displayNameFromDS: name }));
  }

  return {
    meta: options.meta,
    refId: options.refId,
    length: fields[0].values.length,
    fields,
    name,
  };
}

function transformMetricDataToTable(md: MatrixOrVectorResult[], options: TransformOptions): DataFrame {
  if (!md || md.length === 0) {
    return {
      meta: options.meta,
      refId: options.refId,
      length: 0,
      fields: [],
    };
  }

  const valueText = options.responseListLength > 1 || options.valueWithRefId ? `Value #${options.refId}` : 'Value';

  const timeField = getTimeField([]);
  const metricFields = Object.keys(md.reduce((acc, series) => ({ ...acc, ...series.metric }), {}))
    .sort()
    .map((label) => {
      // Labels have string field type, otherwise table tries to figure out the type which can result in unexpected results
      // Only "le" label has a number field type
      const numberField = label === 'le';
      return {
        name: label,
        config: { filterable: true },
        type: numberField ? FieldType.number : FieldType.string,
        values: new ArrayVector(),
      };
    });
  const valueField = getValueField({ data: [], valueName: valueText });

  md.forEach((d) => {
    if (isMatrixData(d)) {
      d.values.forEach((val) => {
        timeField.values.add(val[0] * 1000);
        metricFields.forEach((metricField) => metricField.values.add(getLabelValue(d.metric, metricField.name)));
        valueField.values.add(parseSampleValue(val[1]));
      });
    } else {
      timeField.values.add(d.value[0] * 1000);
      metricFields.forEach((metricField) => metricField.values.add(getLabelValue(d.metric, metricField.name)));
      valueField.values.add(parseSampleValue(d.value[1]));
    }
  });

  return {
    meta: options.meta,
    refId: options.refId,
    length: timeField.values.length,
    fields: [timeField, ...metricFields, valueField],
  };
}

function getLabelValue(metric: PromMetric, label: string): string | number {
  if (metric.hasOwnProperty(label)) {
    if (label === 'le') {
      return parseSampleValue(metric[label]);
    }
    return metric[label];
  }
  return '';
}

function getTimeField(data: PromValue[], isMs = false): MutableField {
  return {
    name: TIME_SERIES_TIME_FIELD_NAME,
    type: FieldType.time,
    config: {},
    values: new ArrayVector<number>(data.map((val) => (isMs ? val[0] : val[0] * 1000))),
  };
}
type ValueFieldOptions = {
  data: PromValue[];
  valueName?: string;
  parseValue?: boolean;
  labels?: Labels;
  displayNameFromDS?: string;
};

function getValueField({
  data,
  valueName = TIME_SERIES_VALUE_FIELD_NAME,
  parseValue = true,
  labels,
  displayNameFromDS,
}: ValueFieldOptions): MutableField {
  return {
    name: valueName,
    type: FieldType.number,
    display: getDisplayProcessor(),
    config: {
      displayNameFromDS,
    },
    labels,
    values: new ArrayVector<number | null>(data.map((val) => (parseValue ? parseSampleValue(val[1]) : val[1]))),
  };
}

function createLabelInfo(labels: { [key: string]: string }, options: TransformOptions) {
  if (options?.legendFormat) {
    const title = renderTemplate(getTemplateSrv().replace(options.legendFormat, options?.scopedVars), labels);
    return { name: title, labels };
  }

  const { __name__, ...labelsWithoutName } = labels;
  const labelPart = formatLabels(labelsWithoutName);
  let title = `${__name__ ?? ''}${labelPart}`;

  if (!title) {
    title = options.query;
  }

  return { name: title, labels: labelsWithoutName };
}

export function getOriginalMetricName(labelData: { [key: string]: string }) {
  const metricName = labelData.__name__ || '';
  delete labelData.__name__;
  const labelPart = Object.entries(labelData)
    .map((label) => `${label[0]}="${label[1]}"`)
    .join(',');
  return `${metricName}{${labelPart}}`;
}

export function renderTemplate(aliasPattern: string, aliasData: { [key: string]: string }) {
  const aliasRegex = /\{\{\s*(.+?)\s*\}\}/g;
  return aliasPattern.replace(aliasRegex, (_match, g1) => {
    if (aliasData[g1]) {
      return aliasData[g1];
    }
    return '';
  });
}

function transformToHistogramOverTime(seriesList: DataFrame[]) {
  /*      t1 = timestamp1, t2 = timestamp2 etc.
            t1  t2  t3          t1  t2  t3
    le10    10  10  0     =>    10  10  0
    le20    20  10  30    =>    10  0   30
    le30    30  10  35    =>    10  0   5
    */
  for (let i = seriesList.length - 1; i > 0; i--) {
    const topSeries = seriesList[i].fields.find((s) => s.name === TIME_SERIES_VALUE_FIELD_NAME);
    const bottomSeries = seriesList[i - 1].fields.find((s) => s.name === TIME_SERIES_VALUE_FIELD_NAME);
    if (!topSeries || !bottomSeries) {
      throw new Error('Prometheus heatmap transform error: data should be a time series');
    }

    for (let j = 0; j < topSeries.values.length; j++) {
      const bottomPoint = bottomSeries.values.get(j) || [0];
      topSeries.values.toArray()[j] -= bottomPoint;
    }
  }

  return seriesList;
}

function sortSeriesByLabel(s1: DataFrame, s2: DataFrame): number {
  let le1, le2;

  try {
    // fail if not integer. might happen with bad queries
    le1 = parseSampleValue(s1.name ?? '');
    le2 = parseSampleValue(s2.name ?? '');
  } catch (err) {
    console.error(err);
    return 0;
  }

  if (le1 > le2) {
    return 1;
  }

  if (le1 < le2) {
    return -1;
  }

  return 0;
}

function parseSampleValue(value: string): number {
  switch (value) {
    case POSITIVE_INFINITY_SAMPLE_VALUE:
      return Number.POSITIVE_INFINITY;
    case NEGATIVE_INFINITY_SAMPLE_VALUE:
      return Number.NEGATIVE_INFINITY;
    default:
      return parseFloat(value);
  }
}