Open Stack

Sorry for being late to the party.  Since we follow mostly DynamoDB, it makes sense not to deviate too much away from DynamoDB’s consistency mode.

>From what I read about DynamoDB, READ consistency is defined to be either strong consistency or eventual consistency.

  "ConsistentRead<http://docs.aws.amazon.com/amazondynamodb/latest/APIReference/API_Query.html#DDB-Query-request-ConsistentRead>": "boolean”,

ConsistentRead<http://docs.aws.amazon.com/amazondynamodb/latest/APIReference/API_Query.html#API_Query_RequestSyntax>

If set to true, then the operation uses strongly consistent reads; otherwise, eventually consistent reads are used.

Strongly consistent reads are not supported on global secondary indexes. If you query a global secondary index with ConsistentRead set to true, you will receive an error message.

Type: Boolean

Required: No

http://docs.aws.amazon.com/amazondynamodb/latest/APIReference/API_Query.html

WRITE consistency is not clearly defined anywhere. From what Werner Vogel’s description, it seems to indicate writes are replicated across availability zones/data centers synchronously. I guess inside data center, writes are replicated asynchronously. And the API doesn’t allow user to specify WRITE consistency level.

http://www.allthingsdistributed.com/2012/01/amazon-dynamodb.html

Considering the above factors and what Cassandra’s capabilities, I propose we use the following model.

READ:

 *   Strong consistency (synchronously replicate to all, maps to Cassandra READ All consistency level)
 *   Eventual consistency (quorum read, maps to Cassandra READ Quorum)
 *   Weak consistency (not in DynamoDB, maps to Cassandra READ ONE)

WRITE:

 *   Strong consistency (synchronously replicate to all, maps to Cassandra WRITE All consistency level)
 *   Eventual consistency (quorum write, maps to Cassandra WRITE Quorum)
 *   Weak consistency (not in DynamoDB, maps to Cassandra WRITE ANY)

For conditional writes (conditional putItem/deletItem), only strong and eventual consistency should be supported.

Thoughts?

Thanks,

Charles

From: Dmitriy Ukhlov <dukhlov at mirantis.com<mailto:dukhlov at mirantis.com>>
Reply-To: "OpenStack Development Mailing List (not for usage questions)" <openstack-dev at lists.openstack.org<mailto:openstack-dev at lists.openstack.org>>
Date: Tuesday, April 29, 2014 at 10:43 AM
To: Illia Khudoshyn <ikhudoshyn at mirantis.com<mailto:ikhudoshyn at mirantis.com>>
Cc: "OpenStack Development Mailing List (not for usage questions)" <openstack-dev at lists.openstack.org<mailto:openstack-dev at lists.openstack.org>>
Subject: Re: [openstack-dev] [MagnetoDB] Configuring consistency draft of concept

Hi Illia,
WEAK/QUORUM instead of true/false it is ok for me.

But we have also STRONG.

What does STRONG mean? In current concept we a using QUORUM and say that it is strong. I guess it is confusing (at least for me) and can have different behavior for different backends.

I believe that from user point of view only 4 usecases exist: write and read with consistency or not.
For example if we use QUORUM for write what is usecase to use read with STRONG one? QUORUM read is enought to get consistent data. Or if we use WEAK (ONE) for consistent write what is the use case to use read from QUORUM? we need to read from ALL.

But we can to use different kinds of backend's abilities to implement consistent and incosistent operation. To provide the best flexibility of backend  specific features I propose to use backend specific configuration section in table schema. In this case you can get much more then in initial concept. For example specify consistensy level ANY instead of ONE for WEAK consistency if you want concentrate on performance of TWO if you want to provide more fault tolerant behavior.

With my proposal we will have only one limitation in comparison with first proposal - We have maximally flexible consistency, but  per table, not per request. We have only 2 choices to specify consistensy per request (true or false). But I believe that it is enough to cover user usecases



On Tue, Apr 29, 2014 at 6:16 AM, Illia Khudoshyn <ikhudoshyn at mirantis.com<mailto:ikhudoshyn at mirantis.com>> wrote:
Hi all,

Dima, I think I understand your reasoning but I have some issues with that. I agree that binary logic is much more straightforward and easy to understand and use. But following that logic, having the only one hardcoded consistency level is even easier and more understandable.
As I can see, the idea of the proposal is to provide user a more fine-grained control on consistency to leverage backend features AND at the same time to not bound ourselves with only this concrete backend's features. In scope of Maksym's proposal choice between WEAK/QUORUM for me is pretty much the same as your FALSE/TRUE. But I'd prefer to have more.

PS Eager to see your new index design


On Tue, Apr 29, 2014 at 7:44 AM, Dmitriy Ukhlov <dukhlov at mirantis.com<mailto:dukhlov at mirantis.com>> wrote:

Hello Maksym,

Thank you for your work!

I suggest you to consider more general approach and hide backend specific staff. I have the next proposal:
1) add support for inconsistent write operation by adding PutItem, UpdateItem and DeleteItem request parameters "consistent" = True of False (as well as GetItem and Query requests)
2) add possibility to set backend specific metadata (it would be nice to use some generic format like json) per table in scope of create table request. I suggest to specify mapping for Cassandra consistency level per operation type (consistent read, inconsistent read, consistent write, inconsistent write)

I agree that now we have a limitation for inconsistent write operation on tables with indexed fields and for requests with specified expected conditions. I have thought about how to overcome this limitation and it seems that I found out solution for index handling without CAS operation. And maybe it is reasonable to redesign it a bit.

On Mon, Apr 28, 2014 at 8:33 AM, MAKSYM IARMAK (CS) <Maksym_Iarmak at symantec.com<mailto:Maksym_Iarmak at symantec.com>> wrote:
Hi,

Because of we can't use inconsistent write if we use indexed table and condition operations which indexes based on (this staff requires the state of data), we have one more issue.

If we want to make write with consistency level ONE (WEAK) to the indexed table, we will have 2 variants:
1. Carry out the operation successfully and implicitly make write to the indexed table with minimally possible consistency level for it (QUORUM);
2. Raise an exception, that we can not perform this operation and list all possible CLs for this operation.

I personally prefer the 2nd variant. So, does anybody have some objections or maybe another ideas?

________________________________
From: MAKSYM IARMAK (CS) [Maksym_Iarmak at symantec.com<mailto:Maksym_Iarmak at symantec.com>]
Sent: Friday, April 25, 2014 9:14 PM
To: openstack-dev at lists.openstack.org<mailto:openstack-dev at lists.openstack.org>
Subject: [openstack-dev] [MagnetoDB] Configuring consistency draft of concept

>So, here is specification draft of concept.

_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev




--
Best regards,
Dmitriy Ukhlov
Mirantis Inc.



--

Best regards,

Illia Khudoshyn,
Software Engineer, Mirantis, Inc.



38, Lenina ave. Kharkov, Ukraine

www.mirantis.com<http://www.mirantis.ru/>

www.mirantis.ru<http://www.mirantis.ru/>



Skype: gluke_work

ikhudoshyn at mirantis.com<mailto:ikhudoshyn at mirantis.com>



--
Best regards,
Dmitriy Ukhlov
Mirantis Inc.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20140430/745b16ae/attachment.html>