Try StarTree Cloud: 30-day free trial

Manage missing data

Most of the time[^1], to create buckets of a given granularity, the data fetching step will query with GROUP BY timeBucket.
Let's consider a granularity of 1 day. What happens if there is no data on a Sunday? There will be no line for this Sunday. There is a great chance this will make things break later. After all, you expect your detection pipeline to know when there is no data. This case can seem non-likely to happen, but as soon as you filter on specific dimensions, the probability of this happening skyrockets.

To manage missing data, use a TimeIndexFiller node. A usual pipeline will look like this:

Here is the corresponding configuration:

{
  "name": "simple-threshold-template",
  "description": "Sample threshold alert. Runs every hour. Change the template properties to run on your data",
  "cron": "0 * * * * ? *",
  "template": {
    "nodes": [
      {
        "name": "root",
        "type": "AnomalyDetector",
        "params": {
          "type": "THRESHOLD",
          "component.monitoringGranularity": "${monitoringGranularity}",
          "component.timestamp": "ts",
          "component.metric": "met",
          "component.max": "${max}",
          "component.min": "${min}",
          "anomaly.metric": "${aggregateFunction}(${metric})"
        },
        "inputs": [
          {
            "targetProperty": "current",
            "sourcePlanNode": "missingDataManager",
            "sourceProperty": "currentData"
          }
        ]
      },
      {
        "name": "missingDataManager",
        "type": "TimeIndexFiller",
        "params": {
          "component.timestamp": "ts"
        },
        "inputs": [
          {
            "sourcePlanNode": "currentDataFetcher",
            "sourceProperty": "currentData"
          }
        ],
        "outputs": [
          {
            "outputName": "currentData"
          }
        ]
      },
      {
        "name": "currentDataFetcher",
        "type": "DataFetcher",
        "params": {
          "component.dataSource": "${dataSource}",
          "component.query": "SELECT __timeGroup(\"${timeColumn}\", '${timeColumnFormat}', '${monitoringGranularity}') as ts, ${aggregateFunction}(${metric}) as met FROM ${dataset} WHERE __timeFilter(\"${timeColumn}\", '${timeColumnFormat}') GROUP BY ts ORDER BY ts LIMIT 1000"
        },
        "outputs": [
          {
            "outputKey": "pinot",
            "outputName": "currentData"
          }
        ]
      }
    ]
  },
  "templateProperties": {
    "dataSource": "pinotQuickStartAzure",
    "dataset": "pageviews",
    "aggregateFunction": "sum",
    "metric": "views",
    "monitoringGranularity": "P1D",
    "timeColumn": "date",
    "timeColumnFormat": "yyyyMMdd",
    "max": "850000",
    "min": "250000"
  }
}

threshold_with_missing_data_management.json

By default, missing buckets are created with a value of 0.

Configuration and behavior

The TimeIndexFiller takes an input and returns it with the time index filled.

  {
    "name": "missingDataManager",
    "type": "TimeIndexFiller",
    "params": {
      "component.timestamp": "ts"
    },
    "inputs": [
      {
        "sourcePlanNode": "currentDataFetcher",
        "sourceProperty": "currentData"
      }
    ],
    "outputs": [
      {
        "outputName": "currentData"
      }
    ]
  }

Notice there is only one parameter.
Because the DataFetcher uses the ThirdEye macros __timeGroup(...) and __timeFilter(...), metadata about the granularity and the time predicate is directly given to the TimeIndexFiller.

Manual configuration

If the input does not use macros, the TimeIndexFiller requires the following parameters:

  • component.monitoringGranularity: the granularity in ISO 8601 format. Eg: P1D.
  • component.metric: the name of the metric column.
  • component.timestamp: the name of the time column.
  • component.minTimeInference: the strategy to infer the minimum time.
  • component.maxTimeInference: the strategy to infer the maximum time.
  • component.lookback: Used when time inference uses a lookback time. In ISO 8601 format.

The possible strategy are:

  • FROM_DATA: the minimum (resp maximum) time is the minimum (resp maximum) time observed in the input. Does not work well if data is missing at the beginning or at the end.
  • FROM_DETECTION_TIME: the minimum (resp maximum) time is the the minimum (resp maximum) of the analysis timeframe
  • FROM_DETECTION_TIME_WITH_LOOKBACK: same as the previous one, with an offset applied of value component.lookback.

See reference documentation.


[^1] For classic SQL databases. Timeseries-focused database often introduce group-by-time-bucket capabilities, with empty buckets materialized. Still these bucketings require a range, or use the first and last value as range, which is not correct in our case.