Proposal: Publishers with handled back pressure from business logic about broadway

[josevalim]

Batchers are about grouping data before "publishing" it to a separate step. So for example, if you want write to RabbitMQ one by one and write to the database ten by ten, you define two batchers.

So 10 different destinationis is totally possible today. 10 processing steps aren't right now but this is tracked in #39.

[josevalim]

Right, you can't do accumulator in Broadway, so if that is a must, Flow is a better option! But we can allow hashing, which will at least guarantee the same processors handle the same data. This can be useful at least for ordering. I have opened #62 for this case. Thanks!

[msaraiva]

Hi @am-kantox. Thanks for your proposal!

We have already discussed some ideas to make Broadway more flexible/pluggable in the near future. However, we decided, for now, to get more feedback like this from the community before we make a decision in that direction.

We certainly going to get back to this issue in the next few days/weeks, so stay tuned and thanks again for your feedback!

[josevalim]

Hi @am-kantox! Thanks for feedback.

I have some questions about your current pipeline.

When you say you have RabbitMQ -> Flow -> RabbitMQ -> Flow, do you mean that:

1. you get the data from RabbitMQ, then you process it with Flow, then you send it to RabbitMQ, and then there is another Flow that gets the data from RabbitMQ again?

2. or you get the data from RabbityMQ, then you process it with Flow, then you send it to RabbitMQ, and continue processing it with Flow (without getting it from RabbitMQ again)?

That will help us steer the discussion in the correct direction.

[am-kantox]

The first one. We try to decouple all Flows as much as possible, hence no ties. Also, No2 might be of interest as well, when the next pipeline is in the same microservice.

…

On Wed, Feb 27, 2019, 6:54 PM José Valim ***@***.***> wrote: Hi @am-kantox <https://github.com/am-kantox>! Thanks for feedback. I have some questions about your current pipeline. When you say you have RabbitMQ -> Flow -> RabbitMQ -> Flow, do you mean that: 1. you get the data from RabbitMQ, then you process it with Flow, then you send it to RabbitMQ, and then there is another Flow that gets the data from RabbitMQ again? 2. or you get the data from RabbityMQ, then you process it with Flow, then you send it to RabbitMQ, and continue processing it with Flow (without getting it from RabbitMQ again)? That will help us steer the discussion in the correct direction. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#50 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AK3INiVj3k-fOVlQ1JUdTYPfHjujs2Bbks5vRsZrgaJpZM4bT0sC> .

[josevalim]

The first one. We try to decouple all Flows as much as possible, hence no ties.

@am-kantox so in this case, you should be able to create multiple broadway pipelines. Correct? Why would publishers be necessary? Or is just that you don't need batching?

[am-kantox]

Eh. Because we basically do not want to think about where we do publish. Slack and usual HTTP endpoint do not like 5K posts per a sec, and I am currently building a back pressure publisher against these endpoints.

…

On Wed, Feb 27, 2019, 7:44 PM José Valim ***@***.***> wrote: The first one. We try to decouple all Flows as much as possible, hence no ties. @am-kantox <https://github.com/am-kantox> so in this case, you should be able to create multiple broadway pipelines. Correct? Why would publishers be necessary? Or is just that you don't need batching? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#50 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AK3INhjpeMQiqsKVSrxeSuyVTbjpgaz0ks5vRtIFgaJpZM4bT0sC> .

[am-kantox]

That does mean that Broadway might take care about number of publishers to Rabbit, or about number of requests per a sec to some other service, assuming we have peak times and quiet times.

[josevalim]

There are two ways we can implement this functionality today with GenStage:

1. You have custom consumers that set the demand to manual to have granular control over the demand

2. You simply block in the consumer until the thing you are calling can accept the request

Almost everyone is doing 2, because it is by far the simplest approach, and that's what Broadway supports. It is also not a major issue to block a consumer.

Have you implemented 1 by any chance? If so, I would love to take a look at it, otherwise, I would postpone this until we have use cases in place.

[am-kantox]

Well, we have not implemented 1 despite I have had a couple of attempts :) Let us postpone this until we have it implemented. Still, the ability to deliberately select a publish method according to the endpoind is valuable IMHO. Like, if there is a DB as a backend, collect batches and do bulk insert. Or report success if N messages were acked from external source.

…

On Wed, Feb 27, 2019, 8:05 PM José Valim ***@***.***> wrote: There are two ways we can implement this functionality today with GenStage: 1. You have custom consumers that set the demand to manual to have granular control over the demand 2. You simply block in the consumer until the thing you are calling can accept the request Almost everyone is doing 2, because it is by far the simplest approach, and that's what Broadway supports. It is also not a major issue to block a consumer. Have you implemented 1 by any chance? If so, I would love to take a look at it, otherwise, I would postpone this until we have use cases in place. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#50 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AK3INta-u-GaDCJCeEW_feSAVRgE19bMks5vRtbtgaJpZM4bT0sC> .

[josevalim]

Still, the ability to deliberately select a publish method according to the endpoind is valuable IMHO.

You can achieve this today by setting up multiple batchers. :)

[am-kantox]

I must be missing something. Butchers are about getting external data, correct me if I am wrong. I am talking about 10 different destinations, after 10 processing steps. Am I missing something?

…

On Wed, Feb 27, 2019, 8:20 PM José Valim ***@***.***> wrote: Still, the ability to deliberately select a publish method according to the endpoind is valuable IMHO. You can achieve this today by setting up multiple batchers. :) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#50 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AK3INqDgGdHgFmDoPelnECIshg2FaSj9ks5vRtqBgaJpZM4bT0sC> .

[am-kantox]

I am not a mind reader :)

Glad to hear this title means exactly this.

[josevalim]

@am-kantox the batchers part though (10 destinations) exists already today and is in the docs. Was that not made clear in the docs? Is there something we could improve?

[am-kantox]

@josevalim I have carefully re-read the docs; it was totally me misunderstanding the pipeline. I was fooled because I have a very biased view based on our existing implementation. I was under an impression all the heavy work is done _in consumers_, after _batching_, not in _processors_. I am positive docs are good enough and other people discovering _Broadway_ would be fine understanding it as it is. The thing is, now I am not convinced by the top of the pyramid on the pipeline. It sounds like _all the messages_ are going through `Broadway.handle_message/3` callback and we are to implement `Flow` there. Is that correct? Even when multiple producers/processors will be allowed, there is no batching / partitioning _before_ `Broadway.handle_message/3` in the pipeline. Yes, we could rely on pattern matching / guarding in it, but that’s not exactly how batching works. So, the whole `Flow` implementation would usually go into the single `handle_message/3` callback, because of `key` option to `Flow.partition/2` is [biased: _in a vast majority of cases_] more complicated than a guard. Is that correct? The above is the major reason I was misunderstanding the batcher’s role.

[josevalim]

If you are using Broadway, I don't see why you would use Flow. We already do all of the partitioning, parallelism, etc for you. Given you said you are doing RabbitMQ -> Flow -> RabbitMQ -> Flow this will be two broadway pipelines: one that gets from RabbitMQ, processes it and sends to another RabbitMQ channel. And another pipeline that gets from RabbitMQ as well and consumes it.

Why would you need both Broadway and Flow?

[am-kantox]

Flow is a more general and powerful abstraction than Broadway that focuses on data as a whole, providing features like aggregation, joins, windows, etc. — README

@josevalim I understand that there are two Broadway pipelines, but this is out of the scope of why Flow, so let’s assume we have one for simplicity.

We have a complicated Flow with several emits and we need to keep and pass an accumulator through the whole Flow. Basically, this is a processing of the huge CSV having several different partitions and a complicated logic to produce different new Ecto schemas out of it. Either I did not get how Broadway could be used instead, or we are stick to Flow for this particular processing. Partitioning inside these Flows is done by, say, properties.

OTOH, we want this Flows to act as plugins agnostic to the real data source. It might be several CSVs to be glued into one, or even many API calls with “a single row” data in each of them. For the latter, we want to use Broadway, processing different data sources and partitioning them, say, by client. This partitioning is relatively easy and does not require Flow.Window and other sophisticated stuff from Flow.

And we want these parts to be independent as much as possible. I hope the above answers the question.