calcite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Julian Hyde <jh...@apache.org>
Subject Re: Materialization performance
Date Wed, 30 Aug 2017 19:31:30 GMT
Consider a “transaction” that involves reads and writes:

  Read from a data structure
  Do some stuff
  Write to the data structure

If steps 2 and 3 depend on what you read in step 1, then you need to prevent anyone from writing
until you have written. A simple CAS won’t solve this. The simplest solution is for the
whole transaction to be in a critical section. It doesn’t really matter whether that is
implemented using an actor or synchronized blocks.

We are mostly in agreement - especially about using immutable data structures for anything
shared between threads. 

Julian


> On Aug 29, 2017, at 2:01 PM, Christian Beikov <christian.beikov@gmail.com> wrote:
> 
> Imagine the holder of the various hash maps is immutable, let's call it "actor". When
a new registration is done, we create a copy of that holder and CAS it. When we query, we
simply get the current value and access it's maps. So MaterializationService could have an
AtomicReference to a holder "actor" just like right now, but we make the maps immutable and
create copies whenever a change occurs. We could hide such details behind a message passing
interface so that remote models can be implemented too, but that seems like a next step.
> 
> The materialization concurrency issues isn't the only problem, what about the general
usage in multithreaded environments? The whole schema is currently bound to a CalciteConnection.
It would be nice if all the context could be shared between multiple connections so that we
avoid having to initialize every connection. Do you have any plans to tackle that or am I
not seeing how to achieve this?
> 
> 
> Mit freundlichen Grüßen,
> ------------------------------------------------------------------------
> *Christian Beikov*
> Am 29.08.2017 um 19:40 schrieb Julian Hyde:
>>> I'd rather have immutable state being CASed(compare-and-swap) to make
>>> the querying cheap and do updates in an optimistic concurrency control manner.
>> Compare and swap only works for one memory address. You can't use it
>> to, say, debit one bank account and credit another.
>> 
>> The set of valid materializations is just about the only mutable state
>> in Calcite and I think it will need to be several interconnected data
>> structures. So, compare-and-swap (or its high-level equivalent,
>> ConcurrentHashMap) won't cut it.
>> 
>> So we could use locks/monitors (the "synchronized" keyword) or we
>> could use an actor. The key difference between the two is who does the
>> work. With a monitor, each customer grabs the key (there is only one
>> key), walks into the bank vault, and moves the money from one deposit
>> box to another. With an actor, there is a bank employee in the vault
>> who is the only person allowed to move money around.
>> 
>> The work done is the same in both models. There are performance
>> advantages of the actor model (the data structures will tend to exist
>> in one core's cache) and there are code simplicity advantages (the
>> critical code is all in one class or package).
>> 
>> The overhead of two puts/gets on an ArrayBlockingQueue per request is
>> negligible. And besides, you can switch to a non-actor implementation
>> of the service if Calcite is single-threaded.
>> 
>> I haven't thought out the details of multi-tenant. It is not true to
>> say that this is "not a primary requirement for
>> the Calcite project." Look at the "data grid (cache)" on the diagram
>> in my "Optiq" talk [1] from 2013. Dynamic materialized views were in
>> from the very start. There can be multiple instances of the actor
>> (each with their own request/response queues), so you could have one
>> per tenant. Also, it is very straightforward to make the actors
>> remote, replacing the queues with RPC over a message broker. Remote
>> actors are called services.
>> 
>> Julian
>> 
>> [1] https://www.slideshare.net/julianhyde/optiq-a-dynamic-data-management-framework
>> 
>> On Tue, Aug 29, 2017 at 8:25 AM, Jesus Camacho Rodriguez
>> <jcamacho@apache.org> wrote:
>>> LGTM, I think by the time we have support for the outer joins, I might have
>>> had time to finish the filter tree index implementation too.
>>> 
>>> -Jesús
>>> 
>>> 
>>> 
>>> On 8/29/17, 3:11 AM, "Christian Beikov" <christian.beikov@gmail.com> wrote:
>>> 
>>>> I'd like to stick to trying to figure out how to support outer joins for
>>>> now and when I have an implementation for that, I'd look into the filter
>>>> tree index if you haven't done it by then.
>>>> 
>>>> 
>>>> Mit freundlichen Grüßen,
>>>> ------------------------------------------------------------------------
>>>> *Christian Beikov*
>>>> Am 28.08.2017 um 20:01 schrieb Jesus Camacho Rodriguez:
>>>>> Christian,
>>>>> 
>>>>> The implementation of the filter tree index is what I was referring to
>>>>> indeed. In the initial implementation I focused on the rewriting coverage,
>>>>> but now that the first part is finished, it is at the top of my list
as
>>>>> I think it is critical to make the whole query rewriting algorithm work
>>>>> at scale. However, I have not started yet.
>>>>> 
>>>>> The filter tree index will help to filter not only based on the tables
used
>>>>> by a given query, but also for queries that do not meet the equivalence
>>>>> classes conditions, filter conditions, etc. We could implement all the
>>>>> preconditions mentioned in the paper, and we could add our own additional
>>>>> ones. I also think that in a second version, we might need to maybe add
>>>>> some kind of ranking/limit as many views might meet the preconditions
for
>>>>> a given query.
>>>>> 
>>>>> It seems you understood how it should work, so if you could help to
>>>>> quickstart that work by maybe implementing a first version of the filter
>>>>> tree index with a couple of basic conditions (table matching and EC matching?),
>>>>> that would be great. I could review any of the contributions you make.
>>>>> 
>>>>> -Jesús
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>> On 8/28/17, 3:22 AM, "Christian Beikov" <christian.beikov@gmail.com>
wrote:
>>>>> 
>>>>>> If the metadata was cached, that would be awesome, especially because
>>>>>> that would also improve the prformance regarding the metadata retrival
>>>>>> for the query currently being planned, although I am not sure how
the
>>>>>> caching would work since the RelNodes are mutable.
>>>>>> 
>>>>>> Have you considered implementing the filter tree index explained
in the
>>>>>> paper? As far as I understood, the whole thing only works when a
>>>>>> redundant table elimination is implemented. Is that the case? If
so, or
>>>>>> if it can be done easily, I'd propose we initialize all the lookup
>>>>>> structures during registration and use them during planning. This
will
>>>>>> improve planning time drastically and essentially handle the scalability
>>>>>> problem you mention.
>>>>>> 
>>>>>> What other MV-related issues are on your personal todo list Jesus?
I
>>>>>> read the paper now and think I can help you in one place or another
if
>>>>>> you want.
>>>>>> 
>>>>>> 
>>>>>> Mit freundlichen Grüßen,
>>>>>> ------------------------------------------------------------------------
>>>>>> *Christian Beikov*
>>>>>> Am 28.08.2017 um 08:13 schrieb Jesus Camacho Rodriguez:
>>>>>>> Hive does not use the Calcite SQL parser, thus we follow a different
path
>>>>>>> and did not experience the problem on the Calcite end. However,
FWIW we
>>>>>>> avoided reparsing the SQL every time a query was being planned
by
>>>>>>> creating/managing our own cache too.
>>>>>>> 
>>>>>>> The metadata providers implement some caching, thus I would expect
that once
>>>>>>> you avoid reparsing every MV, the retrieval time of predicates,
lineage, etc.
>>>>>>> would improve (at least after using the MV for the first time).
However,
>>>>>>> I agree that the information should be inferred when the MV is
loaded.
>>>>>>> In fact, maybe just making some calls to the metadata providers
while the MVs
>>>>>>> are being loaded would do the trick (Julian should confirm this).
>>>>>>> 
>>>>>>> Btw, probably you will find another scalability issue as the
number of MVs
>>>>>>> grows large with the current implementation of the rewriting,
since the´
>>>>>>> pre-filtering implementation in place does not discard many of
the views that
>>>>>>> are not valid to rewrite a given query, and rewriting is attempted
with all
>>>>>>> of them.
>>>>>>> This last bit is work that I would like to tackle shortly, but
I have not
>>>>>>> created the corresponding JIRA yet.
>>>>>>> 
>>>>>>> -Jesús
>>>>>>> 
>>>>>>> 
>>>>>>> 
>>>>>>> 
>>>>>>> On 8/27/17, 10:43 PM, "Rajat Venkatesh" <rvenkatesh@qubole.com>
wrote:
>>>>>>> 
>>>>>>>> Thread Safety and repeated parsing is a problem. We have
experience with
>>>>>>>> managing 10s of materialized views. Repeated parsing takes
more time than
>>>>>>>> execution of the query itself. We also have a similar problem
where
>>>>>>>> concurrent queries (with a different set of materialized
views potentailly)
>>>>>>>> maybe planned at the same time. We solved it through maintaining
a cache
>>>>>>>> and carefully setting the cache in a thread local.
>>>>>>>> Relevant code for inspiration:
>>>>>>>> https://github.com/qubole/quark/blob/master/optimizer/src/main/java/org/apache/calcite/prepare/Materializer.java
>>>>>>>> https://github.com/qubole/quark/blob/master/optimizer/src/main/java/org/apache/calcite/plan/QuarkMaterializeCluster.java
>>>>>>>> 
>>>>>>>> 
>>>>>>>> 
>>>>>>>> On Sun, Aug 27, 2017 at 6:50 PM Christian Beikov <christian.beikov@gmail.com>
>>>>>>>> wrote:
>>>>>>>> 
>>>>>>>>> Hey, I have been looking a bit into how materialized
views perform
>>>>>>>>> during the planning because of a very long test
>>>>>>>>> run(MaterializationTest#testJoinMaterializationUKFK6)
and the current
>>>>>>>>> state is problematic.
>>>>>>>>> 
>>>>>>>>> CalcitePrepareImpl#getMaterializations always reparses
the SQL and down
>>>>>>>>> the line, there is a lot of expensive work(e.g. predicate
and lineage
>>>>>>>>> determination) done during planning that could easily
be pre-calculated
>>>>>>>>> and cached during materialization creation.
>>>>>>>>> 
>>>>>>>>> There is also a bit of a thread safety problem with the
current
>>>>>>>>> implementation. Unless there is a different safety mechanism
that I
>>>>>>>>> don't see, the sharing of the MaterializationService
and thus also the
>>>>>>>>> maps in MaterializationActor via a static instance between
multiple
>>>>>>>>> threads is problematic.
>>>>>>>>> 
>>>>>>>>> Since I mentioned thread safety, how is Calcite supposed
to be used in a
>>>>>>>>> multi-threaded environment? Currently I use a connection
pool that
>>>>>>>>> initializes the schema on new connections, but that is
not really nice.
>>>>>>>>> I suppose caches are also bound to the connection? A
thread safe context
>>>>>>>>> that can be shared between connections would be nice
to avoid all that
>>>>>>>>> repetitive work.
>>>>>>>>> 
>>>>>>>>> Are these known issues which you have thought about how
to fix or should
>>>>>>>>> I log JIRAs for these and fix them to the best of my
knowledge? I'd more
>>>>>>>>> or less keep the service shared but would implement it
using a copy on
>>>>>>>>> write strategy since I'd expect seldom schema changes
after startup.
>>>>>>>>> 
>>>>>>>>> Regarding the repetitive work that partly happens during
planning, I'd
>>>>>>>>> suggest doing that during materialization registration
instead like it
>>>>>>>>> is already mentioned CalcitePrepareImpl#populateMaterializations.
Would
>>>>>>>>> that be ok?
>>>>>>>>> 
>>>>>>>>> --
>>>>>>>>> 
>>>>>>>>> Mit freundlichen Grüßen,
>>>>>>>>> ------------------------------------------------------------------------
>>>>>>>>> *Christian Beikov*
>>>>>>>>> 
> 


Mime
View raw message