%md #Stream-Stream Joins using Structured Streaming (Python) This notebook illustrates different ways of joining streams. We are going to use the the canonical example of ad monetization, where we want to find out which ad impressions led to user clicks. Typically, in such scenarios, there are two streams of data from different sources - ad impressions and ad clicks. Both type of events have a common ad identifier (say, `adId`), and we want to match clicks with impressions based on the `adId`. In addition, each event also has a timestamp, which we will use to specify additional conditions in the query to limit the streaming state.
%md In absence of actual data streams, we are going to generate fake data streams using our built-in "rate stream", that generates data at a given fixed rate.
In absence of actual data streams, we are going to generate fake data streams using our built-in "rate stream", that generates data at a given fixed rate.
Last refresh: Never
from pyspark.sql.functions import rand spark.conf.set("spark.sql.shuffle.partitions", "1") impressions = ( spark .readStream.format("rate").option("rowsPerSecond", "5").option("numPartitions", "1").load() .selectExpr("value AS adId", "timestamp AS impressionTime") ) clicks = ( spark .readStream.format("rate").option("rowsPerSecond", "5").option("numPartitions", "1").load() .where((rand() * 100).cast("integer") < 10) # 10 out of every 100 impressions result in a click .selectExpr("(value - 50) AS adId ", "timestamp AS clickTime") # -50 so that a click with same id as impression is generated later (i.e. delayed data). .where("adId > 0") )
display(impressions)
Showing all 35 rows.
Last refresh: Never
display(clicks)
Showing all 12 rows.
Last refresh: Never
%md Note: - If you get an error saying the join is not supported, the problem may be that you are running this notebook in an older version of Spark. - If you are running on Community Edition, click Cancel above to stop the streams, as you do not have enough cores to run many streams simultaneously.
Note:
- If you get an error saying the join is not supported, the problem may be that you are running this notebook in an older version of Spark.
- If you are running on Community Edition, click Cancel above to stop the streams, as you do not have enough cores to run many streams simultaneously.
Last refresh: Never
%md ### Inner Join Let's join these two data streams. This is exactly the same as joining two batch DataFrames/Datasets by their common key `adId`.
Inner Join
Let's join these two data streams. This is exactly the same as joining two batch DataFrames/Datasets by their common key adId.
Last refresh: Never
display(impressions.join(clicks, "adId"))
Showing all 31 rows.
Last refresh: Never
%md After you start this query, within a minute, you will start getting joined impressions and clicks. The delays of a minute is due to the fact that clicks are being generated with delay over the corresponding impressions. In addition, if you expand the details of the query above, you will find a few timelines of query metrics - the processing rates, the micro-batch durations, and the size of the state. If you keep running this query, you will notice that the state will keep growing in an unbounded manner. This is because the query must buffer all past input as any new input can match with any input from the past.
After you start this query, within a minute, you will start getting joined impressions and clicks. The delays of a minute is due to the fact that clicks are being generated with delay over the corresponding impressions.
In addition, if you expand the details of the query above, you will find a few timelines of query metrics - the processing rates, the micro-batch durations, and the size of the state. If you keep running this query, you will notice that the state will keep growing in an unbounded manner. This is because the query must buffer all past input as any new input can match with any input from the past.
Last refresh: Never
%md ### Inner Join with Watermarking To avoid unbounded state, you have to define additional join conditions such that indefinitely old inputs cannot match with future inputs and therefore can be cleared from the state. In other words, you will have to do the following additional steps in the join. 1. Define watermark delays on both inputs such that the engine knows how delayed the input can be. 1. Define a constraint on event-time across the two inputs such that the engine can figure out when old rows of one input is not going to be required (i.e. will not satisfy the time constraint) for matches with the other input. This constraint can be defined in one of the two ways. a. Time range join conditions (e.g. `...JOIN ON leftTime BETWEN rightTime AND rightTime + INTERVAL 1 HOUR`), b. Join on event-time windows (e.g. `...JOIN ON leftTimeWindow = rightTimeWindow`). Let's apply these steps to our use case. 1. Watermark delays: Say, the impressions and the corresponding clicks can be delayed/late in event-time by at most "10 seconds" and "20 seconds", respectively. This is specified in the query as watermarks delays using `withWatermark`. 1. Event-time range condition: Say, a click can occur within a time range of 0 seconds to 1 minute after the corresponding impression. This is specified in the query as a join condition between `impressionTime` and `clickTime`.
Inner Join with Watermarking
To avoid unbounded state, you have to define additional join conditions such that indefinitely old inputs cannot match with future inputs and therefore can be cleared from the state. In other words, you will have to do the following additional steps in the join.
Define watermark delays on both inputs such that the engine knows how delayed the input can be.
Define a constraint on event-time across the two inputs such that the engine can figure out when old rows of one input is not going to be required (i.e. will not satisfy the time constraint) for matches with the other input. This constraint can be defined in one of the two ways.
a. Time range join conditions (e.g.
...JOIN ON leftTime BETWEN rightTime AND rightTime + INTERVAL 1 HOUR),b. Join on event-time windows (e.g.
...JOIN ON leftTimeWindow = rightTimeWindow).
Let's apply these steps to our use case.
Watermark delays: Say, the impressions and the corresponding clicks can be delayed/late in event-time by at most "10 seconds" and "20 seconds", respectively. This is specified in the query as watermarks delays using
withWatermark.Event-time range condition: Say, a click can occur within a time range of 0 seconds to 1 minute after the corresponding impression. This is specified in the query as a join condition between
impressionTimeandclickTime.
Last refresh: Never
from pyspark.sql.functions import expr # Define watermarks impressionsWithWatermark = impressions \ .selectExpr("adId AS impressionAdId", "impressionTime") \ .withWatermark("impressionTime", "10 seconds ") clicksWithWatermark = clicks \ .selectExpr("adId AS clickAdId", "clickTime") \ .withWatermark("clickTime", "20 seconds") # max 20 seconds late # Inner join with time range conditions display( impressionsWithWatermark.join( clicksWithWatermark, expr(""" clickAdId = impressionAdId AND clickTime >= impressionTime AND clickTime <= impressionTime + interval 1 minutes """ ) ) )
Showing all 41 rows.
Last refresh: Never
%md We are getting the similar results as the previous simple join query. However, if you look at the query metrics now, you will find that after about a couple of minutes of running the query, the size of the state will stabilize as the old buffered events will start getting cleared up.
We are getting the similar results as the previous simple join query. However, if you look at the query metrics now, you will find that after about a couple of minutes of running the query, the size of the state will stabilize as the old buffered events will start getting cleared up.
Last refresh: Never
%md ### Outer Joins with Watermarking Let's extend this use case to illustrate outer joins. Not all ad impressions will lead to clicks and you may want to keep track of impressions that did not produce clicks. This can be done by applying a left outer join on the impressions and clicks. The joined output will not have the matched clicks, but also the unmatched ones (with clicks data being NULL). While the watermark + event-time constraints is optional for inner joins, for left and right outer joins they must be specified. This is because for generating the NULL results in outer join, the engine must know when an input row is not going to match with anything in future. Hence, the watermark + event-time constraints must be specified for generating correct results.
Outer Joins with Watermarking
Let's extend this use case to illustrate outer joins. Not all ad impressions will lead to clicks and you may want to keep track of impressions that did not produce clicks. This can be done by applying a left outer join on the impressions and clicks. The joined output will not have the matched clicks, but also the unmatched ones (with clicks data being NULL).
While the watermark + event-time constraints is optional for inner joins, for left and right outer joins they must be specified. This is because for generating the NULL results in outer join, the engine must know when an input row is not going to match with anything in future. Hence, the watermark + event-time constraints must be specified for generating correct results.
Last refresh: Never
from pyspark.sql.functions import expr # Inner join with time range conditions display( impressionsWithWatermark.join( clicksWithWatermark, expr(""" clickAdId = impressionAdId AND clickTime >= impressionTime AND clickTime <= impressionTime + interval 1 minutes """ ), "leftOuter" ) )
Showing all 163 rows.
Last refresh: Never
%md After starting this query, you will start getting the inner results within a minute. But after a couple of minutes, you will also start getting the outer NULL results.
After starting this query, you will start getting the inner results within a minute. But after a couple of minutes, you will also start getting the outer NULL results.
Last refresh: Never
%md ### Further Information You can read more about stream-stream joins in the following places: - Databricks blog post on stream-stream joins - https://databricks.com/blog/2018/03/13/introducing-stream-stream-joins-in-apache-spark-2-3.html - Apache Programming Guide on Structured Streaming - https://spark.apache.org/docs/latest/structured-streaming-programming-guide.html#stream-stream-joins - Talk at Spark Summit Europe 2017 - https://databricks.com/session/deep-dive-into-stateful-stream-processing-in-structured-streaming
Further Information
You can read more about stream-stream joins in the following places:
- Databricks blog post on stream-stream joins - https://databricks.com/blog/2018/03/13/introducing-stream-stream-joins-in-apache-spark-2-3.html
- Apache Programming Guide on Structured Streaming - https://spark.apache.org/docs/latest/structured-streaming-programming-guide.html#stream-stream-joins
- Talk at Spark Summit Europe 2017 - https://databricks.com/session/deep-dive-into-stateful-stream-processing-in-structured-streaming
Last refresh: Never
Stream-Stream Joins using Structured Streaming (Python)
This notebook illustrates different ways of joining streams.
We are going to use the the canonical example of ad monetization, where we want to find out which ad impressions led to user clicks. Typically, in such scenarios, there are two streams of data from different sources - ad impressions and ad clicks. Both type of events have a common ad identifier (say,
adId), and we want to match clicks with impressions based on theadId. In addition, each event also has a timestamp, which we will use to specify additional conditions in the query to limit the streaming state.Last refresh: Never