Re: eviction files

[Dan Smith]

Hi Roi,

I would probably start from NewLRUClockHand. That's the class that
maintains the actual state of what entries are least recently used.

You can follow the callers of getLRUEntry back to see how it's being used.
One key place is AbstractLRUMap.lruUpdateCallback, which gets called after
updates to the region data. In there, it checks with a
LRUCapacityController to see if something needs to be evicted, and uses
NewLRUClockHand to figure out what to evict.

-Dan

On Thu, Aug 24, 2017 at 9:02 AM, Roi Apelker  wrote:

> Hi,
>
> I am looking into the internals of the eviction process,
>
> Can anyone point me to the most important classes, the main mechanism
> "wheels" etc.?
>
> Thanks,
>
> Roi
>

eviction files

[Roi Apelker]

Hi,

I am looking into the internals of the eviction process,

Can anyone point me to the most important classes, the main mechanism "wheels" 
etc.?

Thanks,

Roi

-Original Message-
From: Roi Apelker 
Sent: Wednesday, August 16, 2017 8:38 PM
To: dev@geode.apache.org
Subject: RE: continuous query internal mechanism questions

It seems like the code in the native client (in the version I have, which may 
be old) send the message to all servers:

CqResultsPtr CqQueryImpl::executeWithInitialResults(uint32_t timeout) {
  ...

  TcrMessage msg(TcrMessage::EXECUTECQ_WITH_IR_MSG_TYPE, m_cqName, 
m_queryString, CqState::RUNNING, isDurable(), m_tccdm);
  TcrMessage reply(true, m_tccdm);
  ChunkedQueryResponse* resultCollector = (new ChunkedQueryResponse(reply));
  reply.setChunkedResultHandler(static_cast(resultCollector));
  reply.setTimeout(timeout);

  GfErrType err = GF_NOERR;
  err = m_tccdm->sendSyncRequest(msg, reply); ..

And sendSyncRequest:
...

for (std::vector::iterator ep = m_endpoints.begin(); ep != 
m_endpoints.end(); ++ep) {
if ((*ep)->connected()) {
  (*ep)->setDM(this);
  opErr = sendRequestToEP(request, reply, *ep);//this will go to 
ThinClientDistributionManager

...


Can this be causing the issue?



-Roi





-Original Message-
From: Jason Huynh [mailto:jasonhu...@apache.org]
Sent: Tuesday, August 15, 2017 9:25 PM
To: dev@geode.apache.org
Subject: Re: continuous query internal mechanism questions

I am not quite sure how native client registers cqs. From my understanding:
with the java api, I believe there is only one message (ExecuteCQ message) that 
is executed on the server side and then replicated to the other nodes through 
the profile (OperationMessage).

It seems the extra ExecuteCQ message failing and then closing the cq might be 
putting the system in a weird state...

On Tue, Aug 15, 2017 at 7:56 AM Roi Apelker  wrote:

> Hi,
>
> I have been examining the continuous query registration mechanism for 
> quite some time This is related to an issue that I have, where 
> sometimes a node crashes (1 node out of 2), and the other one does not 
> send CQ events. The CQ is registered on a partitioned region which 
> resides on these 2 nodes.
>
> I noticed the following behavior, and I wonder if anyone can comment 
> regarding it, if it is justified or not and what is the reason:
>
> 1. When the software using the client (native client) registers for 
> the CQ, a CQ command (ExecuteCQ61) is received on both servers.
>  -- is this normal behaviour? Does the client actually send this 
> command to both servers?
>
> 2. When this command is received by a server, and the CQ is 
> registered, another registration message is sent to the other node via 
> an OperationMessage (REGISTER_CQ)
>  -- it seems that regularly, the server can handle this situation as 
> the second registration identifies the previous one and does not 
> affect it. but the question, why do we need this 2nd registration, if 
> there is a command sent to each server?
>
> 3. For some reason, sometimes there is a failure to complete the first 
> registration (executed by ExecuteCQ61) and then this failure causes a 
> closure to the CQ, which is accompanied with a close request to the 
> other node.
>  -- I assume by now, since 2 registrations and one closure have 
> occurred on node 2, the CQ is still active and the client receives 
> notifications.
>
> 4. Sometimes, 1 out of 5, once node 1 crashes, I get a cleanup 
> operation, caused by the crash (via MemberCrashedEvent), and this also 
> closes the existing CQ, and in this case the CQ in node 2 does not 
> operate anymore and the client receives no notifications.
>  -- fact is, that 4 out of 4 times, I do not get this cleanup by 
> MemberCrashedEvent (maybe due to some other error), and that the CQ 
> notifications are received normally.
>
> Can anyone clear things up for me? Any comment on any of the 
> statements above will be greatly appreciated.
>
> Thanks,
>
> Roi
>
>
> -Original Message-
> From: Roi Apelker
> Sent: Wednesday, August 09, 2017 3:21 PM
> To: dev@geode.apache.org
> Subject: RE: continuous query internal mechanism
>
> Dhanyavad
>
> -Original Message-
> From: Anilkumar Gingade [mailto:aging...@pivotal.io]
> Sent: Tuesday, August 08, 2017 9:55 PM
> To: dev@geode.apache.org
> Subject: Re: continuous query internal mechanism
>
> Registered events, i meant, are events generated for interest 
> registration "region.registerInterest(*)". And CqEvents are for CQs 
> registered.
>
> -Anil.
>
>
> On Tue, Aug 8, 2017 at 12:27 AM, Roi Apelker 
> wrote:
>
> > Shukriya
> >
> > What is the difference between registered events and CQ events?
> >
> > -Original Message-
> > From: Anilkumar Gingade [mailto:aging...@pivotal.io]
> > Sent: Monday, August 07, 2017 10:12 PM
> > To: dev@geode.apache.org
> > Subject: Re: continuous query internal mechanism
> >