Hi Anil,
I agree that it can happen that two threads try to create the index with the
same name with different index expressions concurrently.
This distributed lock will help to avoid this issue, but the same test will
fail.
The first issue is when we have the above case it will create the first index,
but the second(the same name and different expression) will be not created but
the command is successful.
The second issue(these tests which are failing on PR) is that Geode expects
that if run the same command again it will fail as there is already created
index with the same name. With ignoring exceptions it will pass but shouldn't.
So, with a distributed lock we can avoid the issue you mentioned but still has
other issues.
BR,
Mario
________________________________
Šalje: Anilkumar Gingade <aging...@vmware.com>
Poslano: 3. veljače 2022. 16:46
Prima: dev@geode.apache.org <dev@geode.apache.org>
Predmet: Re: Creating index failed
The other problem which exists is; the case where two threads tries to create
index with the same name with different index expression concurrently. I assume
there are ways this could happen.
One solution to address overall issue with index creation on partitioned region
is by taking a distributed lock with the index name. When index creation
request comes, it first acquires a distributed lock with the index name; any
additional index creation with that name will be blocked till the previous
index is created with the same name; during this time if the index creation
comes through local or remote the exception can be ignored. As there is only
one index creation will be in progress for the same request.
-Anil.
On 2/3/22, 4:41 AM, "Mario Kevo" <mario.k...@est.tech> wrote:
Hi devs,
After implementing ignoring exception some tests failed as we allowed now
to pass command again (although it does nothing as the same index is already
created by execution before).
https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fapache%2Fgeode%2Fpull%2F7195&data=04%7C01%7Cagingade%40vmware.com%7C5c8bc7454b9044a05b1308d9e71275dd%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C0%7C0%7C637794888745101984%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=wUnq5WuzXrRnpeq%2FM8Ah1vF3TL8tETKxd%2B35v%2FXUMLg%3D&reserved=0
There is a summary of how it works by now.
When we are creating an index on a partitioned region, the locator sends to
all members to create an index on all data it contains. The partitioned region
is specific as it is normal that you want to index all data which are
distributed on all members. That leads to every member will try to create it
locally and send index create requests to all members on that site.
All members will check if there is an already created index or index
creating is in progress and wait for it. In case a remotely originated request
comes but there is already created index it will respond with Index and send an
acknowledgment to the request sender side. In case it is not created already it
will create an index on that member and then respond to the request sender
side. This behavior is okay if we are using a small number of the server or
using the --member option while creating indexes(which has no sense to use on
the partitioned region as already described down in the mail thread).
The problem is when we are using a larger num of the servers(8 or more) or
just with debugging on. It will slow down the whole process and then can happen
that on some of the servers remotely originated create index request comes
before locally request. In that case, a remotely originated request will see
that there is no index with that name and will create a new one. But the
problem happens after that when a local request comes and there is already
created index it will think that it is from some execution before and throw
IndexNameConflictException.
https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fapache%2Fgeode%2Fblob%2Fdevelop%2Fgeode-core%2Fsrc%2Fmain%2Fjava%2Forg%2Fapache%2Fgeode%2Finternal%2Fcache%2FPartitionedRegion.java%23L8377&data=04%7C01%7Cagingade%40vmware.com%7C5c8bc7454b9044a05b1308d9e71275dd%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C0%7C0%7C637794888745101984%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=0lbpsVQ63FjPaRBvhqmhmAsYp8V0gH3BokmbASBC9hg%3D&reserved=0
The create index command will fail(despite of that the index is created on
all data, some with local requests ad some with remotely originated requests).
There are two problems with this implementation:
1. The user doesn't know that the index is created and will try to
create it again but then it will fail on all servers.
2. The cluster config is updated after the command is finished
successfully, which is correct as we cannot update the cluster config before
anything is done.
The user can use indexes despite that command failed, but the problem is
that after the restart it has nothing in the cluster config and will not create
an index on them.
So the question is what to do in this case? How to avoid this issue?
Ignore exceptions and fix failing tests expect that a new create index
command will pass or disable --member option if partition region is used(or
just document it) and don't send a request on other members as the command will
send to all members to create it. Or maybe something else?
BR,
Mario
________________________________
Šalje: Mario Kevo <mario.k...@est.tech>
Poslano: 14. prosinca 2021. 14:06
Prima: dev@geode.apache.org <dev@geode.apache.org>
Predmet: Odg: Creating index failed
Hi Alexandar,
The cluster config is updated at the end of the command execution, and only
in case, the command is successful.
I created PR with Anlikumar's suggestion, but some tests failed.
https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fapache%2Fgeode%2Fpull%2F7195&data=04%7C01%7Cagingade%40vmware.com%7C5c8bc7454b9044a05b1308d9e71275dd%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C0%7C0%7C637794888745101984%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=wUnq5WuzXrRnpeq%2FM8Ah1vF3TL8tETKxd%2B35v%2FXUMLg%3D&reserved=0
I tried with ignoring exception if it is already created, but in that case,
if run again the create index command with the same name and expression it will
not fail.
BR,
Mario
________________________________
Šalje: Alexander Murmann <amurm...@vmware.com>
Poslano: 7. prosinca 2021. 18:28
Prima: dev@geode.apache.org <dev@geode.apache.org>
Predmet: Re: Creating index failed
Hi Mario!
I agree with you that the user wanted to index all the data in the region
when using a partitioned region. But when the command is not successful, the
cluster config is not updated.
After the server restart, it will not have indexes as it is not stored in
the cluster configuration.
Interesting! If I understand you correctly, the initial request to each
server succeeds, but later ones will fail because the index is already there.
However, the first, successful request should also have updated the cluster
config, right?. Am I misunderstanding something?
________________________________
From: Mario Kevo <mario.k...@est.tech>
Sent: Tuesday, December 7, 2021 06:36
To: dev@geode.apache.org <dev@geode.apache.org>
Subject: Odg: Creating index failed
Hi Jason,
I agree with you that the user wanted to index all the data in the region
when using a partitioned region. But when the command is not successful, the
cluster config is not updated.
After the server restart, it will not have indexes as it is not stored in
the cluster configuration.
So there should be some changes, as the index is created on all members but
the command is not successful.
I'm working on a fix. As soon as possible I will create PR on the already
mentioned ticket.
BR,
Mario
________________________________
Šalje: Jason Huynh <jhu...@vmware.com>
Poslano: 6. prosinca 2021. 18:45
Prima: dev@geode.apache.org <dev@geode.apache.org>
Predmet: Re: Creating index failed
Hi Mario,
A lot of the indexing code pre-dates GFSH. The behavior you are seeing is
when an index is created on a partition region. When creating an index on a
partition region, the idea is that the user wanted to index all the data in the
region. So the server will let all other servers know to create an index on
the partition region.
This is slightly different for an index on a replicated region. That is
when the index can be created on a per member basis, which is what I think the
-member flag is for.
GFSH however defaults to sending the create index message to all members
for any index type from what I remember and from what is being described. That
is why you’ll see the race condition with indexes created on partitioned
regions but the end result being that the index that someone wanted to create
is either created or already there.
-Jason
On 12/6/21, 6:37 AM, "Mario Kevo" <mario.k...@est.tech> wrote:
Hi devs,
While doing some testing, I found the issue which is already reported
there.
https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fissues.apache.org%2Fjira%2Fbrowse%2FGEODE-7875&data=04%7C01%7Cagingade%40vmware.com%7C5c8bc7454b9044a05b1308d9e71275dd%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C0%7C0%7C637794888745101984%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=OPwcdOskUUOSfKao2ydpT4QvFWqLIcJbhWJJ3o%2B8Puk%3D&reserved=0
If we run the create index command it will create an index locally and
send a request to create an index on other members of that region.
The problem happened if the remote request comes before the request
from the locator, in that case, a request from the locator failed with the
following message: Index "index1" already exists. Create failed due to
duplicate name.
This can be reproduced by running 6 servers with DEBUG log level(due to
this system will be slower), creating a partitioned region, and then creating
an index.
Why does the server send remote requests to other members as they will
get a request from the locator to create an index?
Also when running the gfsh command to create an index on one member, it
will send create index requests to all other members. In that case, what is the
purpose of this --member flag?
BR,
Mario
