+-------------------+
| Dev_time|
+-------------------+
|2015-09-18 05:00:20|
|2015-09-18 05:00:21|
|2015-09-18 05:00:22|
|2015-09-18 05:00:23|
|2015-09-18 05:00:24|
|2015-09-18 05:00:25|
|2015-09-18 05:00:26|
|2015-09-18 05:00:27|
|2015-09-18 05:00:37|
|2015-09-18 05:00:37|
|2015-09-18 05:00:37|
|2015-09-18 05:00:38|
|2015-09-18 05:00:39|
+-------------------+
For spark's dataframe, I want to compute the diff of the datetime ,just like in numpy.diff(array)
Generally speaking there is no efficient way to achieve this using Spark DataFrames
. Not to mention things like order become quite tricky in a distributed setup. Theoretically you can use lag
function as follows:
from pyspark.sql.functions import lag, col, unix_timestamp
from pyspark.sql.window import Window
dev_time = (unix_timestamp(col("dev_time")) * 1000).cast("timestamp")
df = sc.parallelize([
("2015-09-18 05:00:20", ), ("2015-09-18 05:00:21", ),
("2015-09-18 05:00:22", ), ("2015-09-18 05:00:23", ),
("2015-09-18 05:00:24", ), ("2015-09-18 05:00:25", ),
("2015-09-18 05:00:26", ), ("2015-09-18 05:00:27", ),
("2015-09-18 05:00:37", ), ("2015-09-18 05:00:37", ),
("2015-09-18 05:00:37", ), ("2015-09-18 05:00:38", ),
("2015-09-18 05:00:39", )
]).toDF(["dev_time"]).withColumn("dev_time", dev_time)
w = Window.orderBy("dev_time")
lag_dev_time = lag("dev_time").over(w).cast("integer")
diff = df.select((col("dev_time").cast("integer") - lag_dev_time).alias("diff"))
## diff.show()
## +----+
## |diff|
## +----+
## |null|
## | 1|
## | 1|
## | 1|
## | 1|
## | 1|
## | 1|
## | 1|
## | 10|
## ...
but it is extremely inefficient (as for window functions move all data to a single partition if no PARTITION BY
clause is provided). In practice it makes more sense to use sliding
method on a RDD (Scala) or implement your own sliding window (Python). See:
- How to transform data with sliding window over time series data in Pyspark
- How to access Spark RDD Array of elements based on index