storm-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Mauro Giusti <mau...@microsoft.com>
Subject Re: Strange time aggregation behavior exhibited by BaseWindowedBolt
Date Wed, 07 Aug 2019 00:14:15 GMT
unsubscribe


Sent from Outlook<http://aka.ms/weboutlook>

________________________________
From: Stig Rohde Døssing <stigdoessing@gmail.com>
Sent: Thursday, August 1, 2019 5:55 AM
To: user@storm.apache.org <user@storm.apache.org>
Subject: Re: Strange time aggregation behavior exhibited by BaseWindowedBolt

> Does the watermark ever gets generated based on actual clock, window interval
No, not when you're using the .withTimestampField method. The watermark gets created on a
set (wall clock time) interval, but the watermark value is based on the timestamp values extracted
from your tuples. The reason this behavior makes sense is that when you use .withTimestampField,
you're essentially saying "use the tuple timestamps to keep track of time, instead of using
real wall clock time to keep track of time".

No, if tuples 9 and 10 are the last tuples in the stream, they never get delivered. The real
wall clock time doesn't matter when you use .withTimestampField. The only way "time passes"
for the watermark generator is if we receive new tuples with newer timestamps. If you want
to use real wall clock time instead, you shouldn't call .withTimestampField. In that case,
the windowing code would be based on real time instead, so 9 and 10 would get delivered as
soon as 3 seconds have passed. The disadvantage to using real time is that processing becomes
a little less predictable when e.g. the machine running the processing is slow, or in case
you're reprocessing old messages.

> that explains why only first few were getting flushed in time window
Great, happy you found the cause.

Den tor. 1. aug. 2019 kl. 14.04 skrev Sandeep Singh <tosandeepsingh@gmail.com<mailto:tosandeepsingh@gmail.com>>:
Thank you very much for the response. Please see my comments inline   sandeep>>

On Thu, Aug 1, 2019 at 5:17 AM Stig Rohde Døssing <stigdoessing@gmail.com<mailto:stigdoessing@gmail.com>>
wrote:
Regarding why you need the 5th tuple, it is happening because you are using timestamp fields.
The windowing code will receive the first 4 tuples and add them to the same window. Until
it receives the 5th tuple, there is no way to tell whether the window is "done", as we might
receive more tuples that fall within the window. The 5th tuple acts as a trigger that tells
the windowing code that the window with the first 4 tuples is now over, and should be delivered
to your bolt.

More specifically, the way it works is that there's a thread running which periodically (every
10 seconds in your case) sets a watermark. The watermark is set to be the timestamp of the
newest received tuple, minus the lag. The watermark is then passed on to a trigger policy,
which decides how to generate windows. The windows are generated from the watermark backwards,
so if e.g. your watermark is 10, your lag is 2 and your interval is 3, it will try to generate
windows for 0-2, 2-5, 5-8. Note that any tuples with timestamp 9 and 10 aren't delivered yet,
as you've said you expect up to 2 seconds of lag, so it isn't safe to close the window containing
them yet. We can't deliver 9 and 10 until we see a tuple with timestamp 10 plus the lag, so
12.

sandeep>> Got that. Does the watermark ever gets generated based on actual clock, window
interval? For above example, will tuple with timestamps 9 and 10 will ever get emitted if
tuple 10 was the last tuple in the stream and then there is no activity say for more tuples
for next 2 minutes?
I think my confusion was that they will eventually flushed out after (window interval + lag).

See https://github.com/apache/storm/blob/21bb1388414d373572779289edc785c7e5aa52aa/storm-client/src/jvm/org/apache/storm/windowing/WaterMarkEventGenerator.java<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fapache%2Fstorm%2Fblob%2F21bb1388414d373572779289edc785c7e5aa52aa%2Fstorm-client%2Fsrc%2Fjvm%2Forg%2Fapache%2Fstorm%2Fwindowing%2FWaterMarkEventGenerator.java&data=02%7C01%7Cmaurgi%40microsoft.com%7C825bc2456a524cb6f1cc08d7167f89e0%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637002609349043209&sdata=edMoP2SqtwVZJVXCknkqozGOcshYfcKFLQYkMfR5Ibk%3D&reserved=0>
and https://github.com/apache/storm/blob/925422a5b5ad1c3329a2c2b44db460ae94f70806/storm-client/src/jvm/org/apache/storm/windowing/WatermarkTimeTriggerPolicy.java<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fapache%2Fstorm%2Fblob%2F925422a5b5ad1c3329a2c2b44db460ae94f70806%2Fstorm-client%2Fsrc%2Fjvm%2Forg%2Fapache%2Fstorm%2Fwindowing%2FWatermarkTimeTriggerPolicy.java&data=02%7C01%7Cmaurgi%40microsoft.com%7C825bc2456a524cb6f1cc08d7167f89e0%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637002609349053210&sdata=nId0RXwbQ4HuZMbRLSA%2BMwITPgxo5bUKrept2j3sKNI%3D&reserved=0>

Regarding why your tuples are getting split, I don't know. Are you maybe running multiple
tasks of the windowing bolt?

sandeep>> Checked the code I was running only 1 task. However I was using Kafka Spout
to receive the messages in my topology. It was possible to get tuples with higher time to
get processed earlier than other. I started sending the messaged in blocking mode (wait for
previous sent to complete). If the watermark and trigger in based on order in which tuples
arrived,  that explains why only first few were getting flushed in time window


Den tir. 30. jul. 2019 kl. 16.11 skrev Sandeep Singh <tosandeepsingh@gmail.com<mailto:tosandeepsingh@gmail.com>>:
Sorry for multiple message with same subject as I had to register with different email address.
To follow up on the thread, can someone explain to me why the tuples with same timestamp are
sometimes sent in two different time windows? And also why sending an extra 5th tuple is required
before storm invokes execute method? Do I need to set a different value for tumbling window
duration or lag?
Thank you for your help in advance
Sandeep

On Mon, Jul 29, 2019 at 7:27 PM Sandeep Singh <tosandeepsingh@gmail.com<mailto:tosandeepsingh@gmail.com>>
wrote:

During testing of my topology which uses Storm's Tumbling window, I see strange behavior how
my stream of tuples are handled and split into different time windows.

I am using a Tumbling window with duration and lag set to 10 seconds:

                val duration = BaseWindowedBolt.Duration.seconds(10)

                myBolt.withTumblingWindow(duration).withTimestampField("timestampField").withLag(duration)



When I send four tuples with timestamp set to same value "now - 1 second" (where now = System.currentTimeMillis()),
I see log messages that storm is able to extract the time information from tuples. However
bolt's "execute(inputWindow: TupleWindow)" method never gets invoked. In my test I wait for
2 minutes. I do not see any log message about late tuples.



When I send five tuples,  the first four with timestamp  "now - 1 second" and last one with
"now + 1 hour", I see Storm is able to extract all the five tuples.  However the execute(inputWindow:
TupleWindow) method is either invoked

  a) only once with first four tuple (the behavior I expected)  or,

  b) twice, first invocation with tuple 1 & 2, second invocation with tuple 3 & 4.
Since all the four tuples have exactly same timestamp, I don't understand why tuples are partitined
in different time windows.

Also the bolt's execute method never get's invoked with 5th tuple. However, sending 5th tuple
(which is well outside the time duration window of 10 seconds) ensure that execute method
is called once or twice for the first four tuples.



Mime
View raw message