storm-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Svend Vanderveken <svend.vanderve...@gmail.com>
Subject Re: Svend's blog - several questions
Date Thu, 06 Feb 2014 08:25:13 GMT
The logic of a map state is to keep a "state" somewhere, you can think of a
Storm state as a big Map of key values, the keys come from the groupBy and
the values are the result of the aggregations. Conceptually, when your
topology is talking to a State, you can imagine it's actually talking to a
big HashMap (only there's a DB behind for persistence + opaque logic for
error handling).

Most of the time, I try not to have any other part of my product that
actually depends on the location or structure the data is stored in DB, so
I do not really need to be super specific about the storage stucture: that
is up to the IBackingMap implementation I am delegating to. Read or write
access to the DB is done via the Storm primitive, not by accessing the DB
directly. Don't forget there's also the stateQuery primitive you can use to
read you stored state from another place.

There are ways to configure column families and column names, have a look
at the super clear storm-cassandra doc to see how to do that with this
implementation: https://github.com/hmsonline/storm-cassandra

My blog post of last year is indeed illustrating a full implementation
including an in-house IBackingMap  implementation, I think that approach is
sometimes needed when we want fine grained control over things. I should
have made more clear that this is not necessarily the default approach to
have.


I hope this makes sense now.

S












On Wed, Feb 5, 2014 at 11:15 PM, Adrian Mocanu <amocanu@verticalscope.com>wrote:

>  Thank you Svend and Adam.
>
>
>
> Svend I'm your reader and that tutorial is very useful. I've been spending
> lots of time looking at the code and that blog post.
>
>
>
> BTW I initially thought you were adding the nulls incorrectly in Q3 below,
> but now I see you're doing it correctly.
>
>
>
> I have a follow up question:
>
> Why do you say that "we do not implement multiget/multiput, we just take
> an existing implementation for Cassandra or Memcached or anything and they
> do what's right for that backend."
>
> I thought that I had to rewrite an IBackingMap implementation to
> correspond to the tuples and schema I have in my database. I use Cassandra.
>
> I started with com.hmsonline.storm.cassandra.trident.CassandraState or
> trident.cassandra.CassandraState (they both implement IBackingMap) and I
> replaced multiGet and multiPut to match my db schema. (well, I'm trying to
> do it)
>
>
>
> You are saying I can use CassandraState as it is? :D
>
> If so how would it even know what table my data should go into? It allows
> you to set the column family and a few other things where state will be
> saved (keyspace, column family, replication, rowKey). By state I think it
> means sth like txID (transaction ID). Do you by any chance know what this
> state that CassandraState is saving is?
>
> So as you can tell I have no idea how to use CassandraState.
>
>
>
> Thanks again!
>
> -A
>
> *From:* Svend Vanderveken [mailto:svend.vanderveken@gmail.com]
> *Sent:* February-05-14 2:56 PM
>
> *To:* user@storm.incubator.apache.org
> *Subject:* Re: Svend's blog - several questions
>
>
>
>
>
>
>
> On Wed, Feb 5, 2014 at 6:22 PM, Adrian Mocanu <amocanu@verticalscope.com>
> wrote:
>
>  I've read Svend's blog [
> http://svendvanderveken.wordpress.com/2013/07/30/scalable-real-time-state-update-with-storm/]
> multiple times and I have a few questions.
>
>
>
>
>
> So you are my reader! Great :D
>
> (you can post your questions on the blog itself, I'm more likely to spot
> it there)
>
>
>
>
>
>
>
> "Because we did a groupBy on one tuple field, each List contains here one
> single
>
> String: the correlationId. Note that the list we return must have exactly
> the same
>
> size as the list of keys, so that Storm knows what period corresponds to
> what key.
>
> So for any key that does not exist in DB, we simply put a null in the
> resulting list."
>
>
>
> Q1: Do the db keys come only from groupBy?
>
>
>
> Yes, the key values arriving in the multiget are the field value by which
> we are grouping
>
> do  groupBy (new Fields("color")) and you get things like "blue"; "green",
> "flowerly romantic red"...
>
>
>
>
>
>
>
>
>
>  Q2: Can you do groupBy multiple keys:like .groupBy("name").groupBy("id")
> ?
>
>
>
> yes, the syntax is like this:
>
>
>
> groupBy (new Fields("name", "id"))
>
>
>
> That's the reason the keys in the multiget are List<Object> and not simply
> Object. We receive them in the order they specified in the topology
> definition
>
>
>
>
>
>  Q3: When we add null we keep the size of the results list the same as
> they keys list but I don't understand how we make sure that key(3) points
> to correct result(3).
>
> After all we're adding nulls at the end of result list not intermitently.
> ie: if
>
> key(1) does not have an entry in db, and key size is 5, we add null to
> last position
>
> in results not to results(1). This doesn't preserve consistency/order so
> key(1) now
>
> gives result(1) which is not null as it should be. Is the code incorrect
> ... or the
>
> explanation on Svend's blog is incorrect?
>
>
>
>
>
> The order should indeed be respected, so if the strategy to handling error
> DB error in a multi-get is to put nulls, that they should indeed be at
> index corresponding to the problematic key. Is there part of my toy project
> code who is padding nulls at the end? If so that's indeed a bug, please let
> me know where (or better, fork and send me a pull request)
>
>
>
> Note that I'm not particularly recommending to put nulls in case of
> unrecoverable errors in a multi-get, that's actually a simplistic way of
> handling the error. The contract with storm is either to fail either to
> return a list of the correct size in the correct order. The data itself and
> its semantic is up to the topology implementation, i.e. up to us.
>
>
>
>
>
>
>
>
>
> Moving on,
>
> "Once this is loaded Storm will present the tuples having the same
> correlation ID
>
> one by one to our reducer, the PeriodBuilder"
>
>
>
> Q4: Does Trident/Storm call the reducer after calling multiGet and before
> calling multiPut?
>
>
>
> yes
>
>
>
>  Q5: What params (and their types) are passed to the reducer and what
> parameters should it emit so they can go into multiGet?
>
>
>
> the reducer is called iteratively, it starts with the state found from DB
> (returned by the multiget) and the first grouped tuple, then the second,
> then the third... until the last tuple. The return value of the last call
> of the reducer is what is provided to the multiput, for the same key as the
> multiget.
>
>
>
> "reduce" is actually a very common pattern in functional programming,
> which us java programming are sometimes less aware of. Look up some general
> doc on "reduce", the storm approach to it is very traditional, i.e. Storm
> has defined the "reduce" primitive exactly the way many other tools are
> defining that primitive
>
>
>
>
>
> Q6: The first time the program is run the database is empty and multiGet
> will return nothing.
>
> Does the reducer need to take care and make sure to insert for the first
> time as opposed to update value? I do see that reducer (TimelineUpdater)
> checks for nulls and I'm guessing this is the reason why it does so.
>
>
>
>
>
>
>
> Exactly.
>
>
>
> That's also why returning null in case of error in the multiget is
> questionable and probably not what you would systematically do: it is
> equivalent to saying: there's garbage in persistence for that key, so let's
> just consider there's nothing. The actually proper thing to do depends on
> the task at hand, but actually, such error in multiget is ofter a symptom
> that we stored garbage in persistence in the past due to some other, it's
> too late to correct it now.
>
>
>
> Last thing: most of the time we do not implement multiget/multiput, we
> just take an existing implementation for Cassandra or Memcached or anything
> and they do what's right for that backend.
>
>
>
>
>
>  Q7:
>
> Can someone explain what these mean:
>
> .each  (I've seen this used even consecutively: .each(..).each(..) )
>
> .newStream
>
> .newValuesStream
>
> .persistAggregate
>
>
>
>
>
> I think they are all detailed here:
> https://github.com/nathanmarz/storm/wiki/Trident-API-Overview
>
>
>
>
>
> I am unable to find javadocs with documentation for the method signatures.
>
> These java docs don't help much:
> http://nathanmarz.github.io/storm/doc/storm/trident/Stream.html
>
>
>
>
>
> Q8:
>
> Storm has ack/fail; does Trident handle that automatically?
>
>
>
>
>
> Yes, although you can also explicitly trigger error. Look up my next blog:
>  error handling in Storm Trident.
>
>
>
>
>
>
>
>
>
> Q9: Has anyone tried Spark? http://spark.incubator.apache.org/streaming/
>
> I'm wondering if anyone has tried it because I'm thinking of ditching
> storm and moving to that.
>
> It seems much much much better documented.
>
>
>
>
>
> Spark looks cool  I've not played with it yet, no. Go ahead, keep us
> posted what you find out!
>
>
>
>
>
>
>
> Lots of questions I know. Thanks for reading!
>
>
>
>
>
> and you :D
>
>
>
>
>
> -Adrian
>
>
>
>
>
>
>
> Svend
>

Mime
View raw message