在 pyspark 中创建一个带有时间戳和 groupby 的 KPI

Question

I have a dataframe containing logs just like this example:

+------------+--------------------------+--------------------+-------------------+
|Source      |Error                     |          @timestamp| timestamp_rounded |
+------------+--------------------------+--------------------+-------------------+
|      A     |             No           |2021-09-12T14:07:...|2021-09-12 16:10:00|
|      B     |             No           |2021-09-12T12:49:...|2021-09-12 14:50:00|
|      C     |             No           |2021-09-12T12:59:...|2021-09-12 15:00:00|
|      C     |             No           |2021-09-12T12:58:...|2021-09-12 15:00:00|
|      B     |             No           |2021-09-12T14:22:...|2021-09-12 16:20:00|
|      A     |             Yes          |2021-09-12T14:22:...|2021-09-12 16:25:00|
|      B     |             No           |2021-09-12T13:00:...|2021-09-12 15:00:00|
|      B     |             No           |2021-09-12T12:57:...|2021-09-12 14:55:00|
|      B     |             No           |2021-09-12T12:57:...|2021-09-12 15:00:00|
|      B     |             No           |2021-09-12T12:58:...|2021-09-12 15:00:00|
|      C     |             No           |2021-09-12T12:54:...|2021-09-12 14:55:00|
|      A     |             Yes          |2021-09-12T14:17:...|2021-09-12 16:15:00|
|      B     |             No           |2021-09-12T12:43:...|2021-09-12 14:45:00|
|      A     |             No           |2021-09-12T12:45:...|2021-09-12 14:45:00|
|      D     |             No           |2021-09-12T12:57:...|2021-09-12 14:55:00|
|      A     |             No           |2021-09-12T13:00:...|2021-09-12 15:00:00|
|      C     |             No           |2021-09-12T12:47:...|2021-09-12 14:45:00|
|      A     |             No           |2021-09-12T12:57:...|2021-09-12 15:00:00|
|      A     |             No           |2021-09-12T13:00:...|2021-09-12 15:00:00|
|      A     |             No           |2021-09-12T14:23:...|2021-09-12 16:25:00|
+------------+--------------------------+--------------------+-------------------+
only showing top 20 rows

My dataframe has million of logs, not that it matters.

I would like to calculate the error rate of every source, for every 5 minutes . I have searched for documentation on transformations like this one (groupby with partition? double groupby?...) but I haven't found a lot of information.

I can get a new column with Yes ==> 1 and No ==> 0 and then get the mean for every source with gorupby and {avg: foo} to get the error rate for every source, but I want it to be for every 5 min (see col 'timestamp_rounded')

The result would be like:

+-------------------+------------+--------------+-------------+------------+
|timestamp_rounded  |Error_rate_A| Error_rate_B | Error_rate_C|Error_rate_D|
+-------------------+------------+--------------+-------------+------------+
|2021-09-12 16:10:00|       0    |       0.2    |       0     |       0.2  |
|2021-09-12 16:15:00|       0.1  |       0.3    |       0     |       0    |
|2021-09-12 16:20:00|       0    |       0.2    |       0     |       0    |
|2021-09-12 16:25:00|       0    |       0.2    |       0     |       0    |
|2021-09-12 16:30:00|       0    |       0.2    |       0     |       0    |
|2021-09-12 16:35:00|       0.2  |       0.2    |       0     |       0    |
|2021-09-12 16:40:00|       0.3  |       0.2    |       0     |       0.2  |
|2021-09-12 16:45:00|       0.4  |       0.3    |       0     |       0    |

etc...

Sources can be very numerous (my example has 4 but there can be thousands of sources)

Please tell me if you need more information. Thanks a lot !

Answer 1

Assuming your data is accessible in a dataframe named logs you could achieve this with an initial group by on timestamp_rounded then a pivot on source to transpose your aggregated error rates to rows with columns for each source error rate for each timestamp_rounded . Finally, you may replace missing error rate values with 0.0

Before performing these transformations, we can transform your Yes / No values to 1 / 0 to simplify the aggregation/mean and rename the source column values with a prefix Error_rate_ to achieve the desired column names after the pivot.

NB. I changed 1 of your records in the sample data in the question

|      A     |             No           |2021-09-12T12:57:...|2021-09-12 15:00:00|

to

|      A     |             Yes           |2021-09-12T12:57:...|2021-09-12 15:00:00|

to receive more variation in your data. As a result your dataframe would look like this after your initial aggregation.

You may achieve this using the following:

output_df =(
    logs.withColumn("Error",F.when(F.col("Error")=="Yes",1).otherwise(0))
        .withColumn("Source",F.concat(F.lit("Error_rate_"),F.col("Source")))
        .groupBy("timestamp_rounded")
        .pivot("Source")
        .agg(
            F.round(F.mean("Error"),2).alias("Error_rate")
        )
        .na.fill(0.0)
)

Outputs

+-------------------+------------+------------+------------+------------+
|timestamp_rounded  |Error_rate_A|Error_rate_B|Error_rate_C|Error_rate_D|
+-------------------+------------+------------+------------+------------+
|2021-09-12 14:50:00|0.0         |0.0         |0.0         |0.0         |
|2021-09-12 16:15:00|1.0         |0.0         |0.0         |0.0         |
|2021-09-12 16:20:00|0.0         |0.0         |0.0         |0.0         |
|2021-09-12 16:25:00|0.5         |0.0         |0.0         |0.0         |
|2021-09-12 14:55:00|0.0         |0.0         |0.0         |0.0         |
|2021-09-12 14:45:00|0.0         |0.0         |0.0         |0.0         |
|2021-09-12 16:10:00|0.0         |0.0         |0.0         |0.0         |
|2021-09-12 15:00:00|0.33        |0.0         |0.0         |0.0         |
+-------------------+------------+------------+------------+------------+

NB. The output above is not ordered and can easily be ordered using .orderBy

Let me know if this works for you.