Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation
On 09/02/2014 07:15 AM, Duncan Thomas wrote: On 11 August 2014 19:26, Jay Pipes jaypi...@gmail.com wrote: The above does not really make sense for MySQL Galera/PXC clusters *if only Galera nodes are used in the cluster*. Since Galera is synchronously replicated, there's no real point in segregating writers from readers, IMO. Better to just spread the write AND read load equally among all Galera cluster nodes. Unfortunately it is possible to get bitten by the difference between 'synchronous' and 'virtually synchronous' in practice. Not in my experience. The thing that has bitten me in practice are Galera's lack of support for SELECT FOR UPDATE, which is used extensively in some of the OpenStack projects. Instead of taking a write-intent lock on one or more record gaps (which is what InnoDB does in the case of a SELECT FOR UPDATE on a local node), Galera happily replicates DML statements to all other nodes in the cluster. If two of those nodes attempt to modify the same row or rows in a table, then the working set replication will fail to certify, which results in a certification timeout, which is then converted to an InnoDB deadlock error. It's the difference between hanging around waiting on a local node for the transaction that called SELECT FOR UPDATE to complete and release the write-intent locks on a set of table rows versus hanging around waiting for the InnoDB deadlock/lock timeout to bubble up from the working set replication certification (which typically is longer than the time taken to lock the rows in a single transaction, and therefore causes thundering herd issues with the conductor attempting to retry stuff due to the use of the @retry_on_deadlock decorator which is so commonly used everywhere) FWIW, I've cc'd a real expert on the matter. Peter, feel free to clarify, contradict, or just ignore me :) Best, -jay ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation
On 11 August 2014 19:26, Jay Pipes jaypi...@gmail.com wrote: The above does not really make sense for MySQL Galera/PXC clusters *if only Galera nodes are used in the cluster*. Since Galera is synchronously replicated, there's no real point in segregating writers from readers, IMO. Better to just spread the write AND read load equally among all Galera cluster nodes. Unfortunately it is possible to get bitten by the difference between 'synchronous' and 'virtually synchronous' in practice. I'll try to get somebody more knowledgeable about the details to comment further. ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation
Lee, No problem about mixing up the Mike's, there's a bunch of us out there :-). What are you are describing here is very much like a spec I wrote for Nova[1] a couple months ago and then never got back to. At the time I considered gearing the feature toward oslo.db and I can't remember exactly why I didn't. I think it probably had more to do with having folks that are familiar with the problem reviewing code in Nova than anything else. Anyway, I'd like to revisit this in Kilo or if you see a nice way to integrate this into oslo.db I'd love to see your proposal. -Mike [1] https://review.openstack.org/#/c/93466/ On Sun, Aug 10, 2014 at 10:30 PM, Li Ma skywalker.n...@gmail.com wrote: not sure if I said that :). I know extremely little about galera. Hi Mike Bayer, I'm so sorry I mistake you from Mike Wilson in the last post. :-) Also, say sorry to Mike Wilson. I’d totally guess that Galera would need to first have SELECTs come from a slave node, then the moment it sees any kind of DML / writing, it transparently switches the rest of the transaction over to a writer node. You are totally right. @transaction.writer def read_and_write_something(arg1, arg2, …): # … @transaction.reader def only_read_something(arg1, arg2, …): # … The first approach that I had in mind is the decorator-based method to separates read/write ops like what you said. To some degree, it is almost the same app-level approach to the master/slave configuration, due to transparency to developers. However, as I stated before, the current approach is merely used in OpenStack. Decorator is more friendly than use_slave_flag or something like that. If ideally transparency cannot be achieved, to say the least, decorator-based app-level switching is a great improvement, compared with the current implementation. OK so Galera would perhaps have some way to make this happen, and that's great. If any Galera expert here, please correct me. At least in my experiment, transactions work in that way. this (the word “integrate”, and what does that mean) is really the only thing making me nervous. Mike, just feel free. What I'd like to do is to add a django-style routing method as a plus in oslo.db, like: [database] # Original master/slave configuration master_connection = slave_connection = # Only Support Synchronous Replication enable_auto_routing = True [db_cluster] master_connection = master_connection = ... slave_connection = slave_connection = ... HOWEVER, I think it needs more investigation, so this is why I'd like to put it in the mailing list in the early stage to raise some discussions in depth. I'm not a Galera expert. I really appreciate any challenges here. Thanks, Li Ma - Original Message - From: Mike Bayer mba...@redhat.com To: OpenStack Development Mailing List (not for usage questions) openstack-dev@lists.openstack.org Sent: 星期日, 2014年 8 月 10日 下午 11:57:47 Subject: Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation On Aug 10, 2014, at 11:17 AM, Li Ma skywalker.n...@gmail.com wrote: How about Galera multi-master cluster? As Mike Bayer said, it is virtually synchronous by default. It is still possible that outdated rows are queried that make results not stable. not sure if I said that :). I know extremely little about galera. Let's move forward to synchronous replication, like Galera with causal-reads on. The dominant advantage is that it has consistent relational dataset support. The disadvantage are that it uses optimistic locking and its performance sucks (also said by Mike Bayer :-). For optimistic locking problem, I think it can be dealt with by retry-on-deadlock. It's not the topic here. I *really* don’t think I said that, because I like optimistic locking, and I’ve never used Galera ;). Where I am ignorant here is of what exactly occurs if you write some rows within a transaction with Galera, then do some reads in that same transaction. I’d totally guess that Galera would need to first have SELECTs come from a slave node, then the moment it sees any kind of DML / writing, it transparently switches the rest of the transaction over to a writer node. No idea, but it has to be something like that? So, the transparent read/write separation is dependent on such an environment. SQLalchemy tutorial provides code sample for it [1]. Besides, Mike Bayer also provides a blog post for it [2]. So this thing with the “django-style routers”, the way that example is, it actually would work poorly with a Session that is not in “autocommit” mode, assuming you’re working with regular old databases that are doing some simple behind-the-scenes replication. Because again, if you do a flush, those rows go to the master, if the transaction is still open, then reading from the slaves you won’t see the rows you just inserted.So in reality, that example is kind of crappy
Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation
Hi Li, comments inline. On 08/08/2014 12:03 AM, Li Ma wrote: Getting a massive amount of information from data storage to be displayed is where most of the activity happens in OpenStack. The two activities of reading data and writing (creating, updating and deleting) data are fundamentally different. The optimization for these two opposite database activities can be done by physically separating the databases that service these two different activities. All the writes go to database servers, which then replicates the written data to the database server(s) dedicated to servicing the reads. Currently, AFAIK, many OpenStack deployment in production try to take advantage of MySQL (includes Percona or MariaDB) multi-master Galera cluster. It is possible to design and implement a read/write separation schema for such a DB cluster. The above does not really make sense for MySQL Galera/PXC clusters *if only Galera nodes are used in the cluster*. Since Galera is synchronously replicated, there's no real point in segregating writers from readers, IMO. Better to just spread the write AND read load equally among all Galera cluster nodes. However, if you have a Galera cluster that then slaves off to one or more standard MySQL slaves, then certainly doing writer/reader segregation could be useful, especially for directing readers of aggregate or report-type data to the read-only slaves. Actually, OpenStack has a method for read scalability via defining master_connection and slave_connection in configuration, but this method lacks of flexibility due to deciding master or slave in the logical context(code). It's not transparent for application developer. As a result, it is not widely used in all the OpenStack projects. So, I'd like to propose a transparent read/write separation method for oslo.db that every project may happily takes advantage of it without any code modification. I've never seen a writer/reader segregation proxy or middleware piece that was properly able to send the right reads to the slaves. Unfortunately, determining what are the right reads to send to the slaves is highly application-dependent, since the application knows when it can tolerate slave lags. Moreover, I'd like to put it in the mailing list in advance to make sure it is acceptable for oslo.db. I think oslo.db is not the right place for this. I believe the efforts that Mike Wilson has been doing in the slavification blueprints are the more appropriate place to add this slave-aware code. Best, -jay I'd appreciate any comments. br. Li Ma ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation
Li Ma, Mike [Wilson | Bayer], and Roman Podoliaka, A similar topic came up in Atlanta at a database panel I participated in. Jay Pipes had organized it as part of the ops track and Peter Boros (of Percona) and I were on the panel. The issue of what to do about the database under OpenStack in the face of high load from such components as, for example ceilometer. Splitting reads and writes is a solution that is fraught with challenges as it requires the application to know where it wrote, where it should read from, what is replication latency, and all of that. At the heart of the issue is that you want to scale the database. I had suggested at this panel that those who want to try and solve this problem should try the Database Virtualization Engine[1] product from Tesora. In the interest of full disclosure, I work for Tesora. The solution is a simple way to horizontally scale a MySQL (or Percona or MariaDB) database across a collection of database servers. It exposes a MySQL compatible interface and takes care of all the minutiae of where to store data, partitioning it across the various database servers, and executing queries on behalf of an application irrespective of the location of the data. It natively provides such capabilities as distributed joins, aggregation and sorting. Architecturally it is a traditional parallel database built from a collection of unmodified MySQL (or variant) databases. It is open source, and available for free download.[2] Percona recently tested[3] the DVE product and confirmed that the solution provided horizontal scalability and linear (and in some cases better than linear) performance improvements[4] with scale. You can get a copy of their full test report here.[5] Ingesting data at very high volume is often an area of considerable pain for large systems and in one demonstration of our product, we were required to ingest 1 million CDR style records per second. We demonstrated that with just 15 Amazon RDS servers (m1.xlarge, standard EBS volumes, no provisioned IOPS) and two c1.xlarge servers to run the Tesora DVE software, we could in fact ingest a sustained stream of over 1 million CDR’s a second![6] To Mike Wilson and Roman’s point, the solution I’m proposing would be entirely transparent to the developer and would be something that would be both elastic and scalable with the workload placed on it. In addition, standard SQL queries will continue to work unmodified, irrespective of which database server physically holds a row of data. To Mike Bayer’s point about data distribution and transaction management; yes, we handle all the details relating to handling data consistency and providing atomic transactions during Insert/Update/Delete operations. As a company, we at Tesora are committed to OpenStack and are significant participants in Trove (the database-as-a-service project for OpenStack). You can verify this yourself on Stackalytics [7] or [8]. If you would like to consider it as a part of your solution to oslo.db, we’d be thrilled to work with the OpenStack community to make this work, both from a technical and a business/licensing perspective. You can catch most of our dev team on either #openstack-trove or #tesora. Some of us from Tesora, Percona and Mirantis are planning an ops panel similar to the one at Atlanta, for the Summit in Paris. I would definitely like to meet with more of you in Paris and discuss how we address issues of scale in the database that powers an OpenStack implementation. Thanks, -amrith -- Amrith Kumar, CTO Tesora (www.tesora.com) Twitter: @amrithkumar IRC: amrith @freenode [1] http://www.tesora.com/solutions/database-virtualization-engine [2] http://www.tesora.com/solutions/downloads/products [3] http://www.mysqlperformanceblog.com/2014/06/24/benchmarking-tesoras-database-virtualisation-engine-sysbench/ [4] http://www.tesora.com/blog/perconas-evaluation-our-database-virtualization-engine [5] http://resources.tesora.com/site/download/percona-benchmark-whitepaper [6] http://www.tesora.com/blog/ingesting-over-100-rows-second-mysql-aws-cloud [7] http://stackalytics.com/?module=trove-group http://stackalytics.com/?module=trove-groupmetric=commits metric=commits [8] http://stackalytics.com/?module=trove-group http://stackalytics.com/?module=trove-groupmetric=marks metric=marks From: Mike Wilson [mailto:geekinu...@gmail.com] Sent: Friday, August 08, 2014 7:35 PM To: OpenStack Development Mailing List (not for usage questions) Subject: Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation Li Ma, This is interesting, In general I am in favor of expanding the scope of any read/write separation capabilities that we have. I'm not clear what exactly you are proposing, hopefully you can answer some of my questions inline. The thing I had thought of immediately was detection
Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation
On Aug 10, 2014, at 9:59 AM, Amrith Kumar amr...@tesora.com wrote: To Mike Bayer’s point about data distribution and transaction management; yes, we handle all the details relating to handling data consistency and providing atomic transactions during Insert/Update/Delete operations. As a company, we at Tesora are committed to OpenStack and are significant participants in Trove (the database-as-a-service project for OpenStack). You can verify this yourself on Stackalytics [7] or [8]. If you would like to consider it as a part of your solution to oslo.db, we’d be thrilled to work with the OpenStack community to make this work, both from a technical and a business/licensing perspective. You can catch most of our dev team on either #openstack-trove or #tesora. Some of us from Tesora, Percona and Mirantis are planning an ops panel similar to the one at Atlanta, for the Summit in Paris. I would definitely like to meet with more of you in Paris and discuss how we address issues of scale in the database that powers an OpenStack implementation. OK well just to be clear, oslo.db is Python code that basically provides in-application helpers and patterns to work with databases, primarily through SQLAlchemy. So it’s essentially openstack-specific patterns and recipes on top of SQLAlchemy. It has very little to do with the use of special database backends that know how to partition among shards and/or master/slaves (I thought the original proposal was for master/slave).So the Tesora product would be 99% “drop in”, with at most some configurational flags set up on the Python side, and everything else being configurational. Since the proposal here is for “transparent”, which is taken to mean, “no app changes are needed”. My only point was that, an application-layer reader/writer distribution approach would need to work at the level of transactions, not statements, and therefore would need to know at transaction start time what the nature of the transaction would be (and thus requires some small declaration at the top, hence code changes…code changes that I think are a good thing as explicit declaration of reader/writer methods up top can be handy in other ways too). Thanks, -amrith -- Amrith Kumar, CTO Tesora (www.tesora.com) Twitter: @amrithkumar IRC: amrith @freenode [1] http://www.tesora.com/solutions/database-virtualization-engine [2] http://www.tesora.com/solutions/downloads/products [3] http://www.mysqlperformanceblog.com/2014/06/24/benchmarking-tesoras-database-virtualisation-engine-sysbench/ [4] http://www.tesora.com/blog/perconas-evaluation-our-database-virtualization-engine [5] http://resources.tesora.com/site/download/percona-benchmark-whitepaper [6] http://www.tesora.com/blog/ingesting-over-100-rows-second-mysql-aws-cloud [7] http://stackalytics.com/?module=trove-groupmetric=commits [8] http://stackalytics.com/?module=trove-groupmetric=marks From: Mike Wilson [mailto:geekinu...@gmail.com] Sent: Friday, August 08, 2014 7:35 PM To: OpenStack Development Mailing List (not for usage questions) Subject: Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation Li Ma, This is interesting, In general I am in favor of expanding the scope of any read/write separation capabilities that we have. I'm not clear what exactly you are proposing, hopefully you can answer some of my questions inline. The thing I had thought of immediately was detection of whether an operation is read or write and integrating that into oslo.db or sqlalchemy. Mike Bayer has some thoughts on that[1] and there are other approaches around that can be copied/learned from. These sorts of things are clear to me and while moving towards more transparency for the developer, still require context. Please, share with us more details on your proposal. -Mike [1] http://www.percona.com/doc/percona-xtradb-cluster/5.5/wsrep-system-index.html [2] http://techspot.zzzeek.org/2012/01/11/django-style-database-routers-in-sqlalchemy/ On Thu, Aug 7, 2014 at 10:03 PM, Li Ma skywalker.n...@gmail.com wrote: Getting a massive amount of information from data storage to be displayed is where most of the activity happens in OpenStack. The two activities of reading data and writing (creating, updating and deleting) data are fundamentally different. The optimization for these two opposite database activities can be done by physically separating the databases that service these two different activities. All the writes go to database servers, which then replicates the written data to the database server(s) dedicated to servicing the reads. Currently, AFAIK, many OpenStack deployment in production try to take advantage of MySQL (includes Percona or MariaDB) multi-master Galera cluster. It is possible to design and implement a read/write separation schema for such a DB cluster
Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation
Thanks for all the detailed analysis, Mike W, Mike B, and Roman. For a production-ready database system, replication is a must I think. So, the questions are which replication mode is suitable for OpenStack and which way is suitable for OpenStack to improve performance and scalability of DB access. In current implementation of database API in OpenStack, master/slave connection is defined for optimizing the performance. Developers of each OpenStack component take the responsibility of making use of it in the application context and some other guys take the responsibility of architecting database system to meet the requirements in various production environments. No general guideline for it. Actually, it is not that easy to determine which transaction is able to be conducted by slave due to data consistency and business logic for different OpenStack components. The current status is that master/slave configuration is not widely used and only Nova uses slave connection in its periodic tasks which are not sensitive to the status of replication. Due to the nature of asynchronous replication, query to DB is not stable, so the risks of using slaves are apparent. How about Galera multi-master cluster? As Mike Bayer said, it is virtually synchronous by default. It is still possible that outdated rows are queried that make results not stable. When using such eventual consistency methods, you have to carefully design which transaction is tolerant of old data. AFAIK, no matter which component is, Nova, Cinder or Neutron, most of the transactions are not that 'tolerant'. As Mike Bayer said, consistent relational dataset is very important. As a footnote, consistent relational dataset is very important for OpenStack components. This is why only non-sensitive periodic tasks are using slaves in Nova. Let's move forward to synchronous replication, like Galera with causal-reads on. The dominant advantage is that it has consistent relational dataset support. The disadvantage are that it uses optimistic locking and its performance sucks (also said by Mike Bayer :-). For optimistic locking problem, I think it can be dealt with by retry-on-deadlock. It's not the topic here. If we first ignore the performance-suck problem, multi-master cluster with synchronous replication is the perfect for OpenStack with any masters+slaves enabled and it can truly scale-out. So, the transparent read/write separation is dependent on such an environment. SQLalchemy tutorial provides code sample for it [1]. Besides, Mike Bayer also provides a blog post for it [2]. What I did is to re-implement it in OpenStack DB API modules in my development environment, using Galera cluster(causal-reads on). It has been running perfectly for more than a week. The routing session manager works well while maintaining data consistency. Back to the performance-suck problem, theoretically causal-reads-on will definitely affect the overall performance of concurrent DB reads, but I cannot find any report(officially or unofficially) on causal-reads-performance-degradation. Actually in the production system of my company, the Galera performance is tuned via network round-trip time, network throughput, number of slave threads, keep-alive and wsrep flow control parameters. All in all, firstly, transparent read/write separation is feasible using synchronous replication method. Secondly, it may help scale-out in large deployment without any code modification. Moreover, it needs fine-tuning (Of course, every production system needs it :-). Finally, I think if we can integrate it into oslo.db, it is a perfect plus for those who would like to deploy Galera (or other similar technology) as DB backend. [1] http://docs.sqlalchemy.org/en/rel_0_9/orm/session.html#custom-vertical-partitioning [2] http://techspot.zzzeek.org/2012/01/11/django-style-database-routers-in-sqlalchemy/ [3] Galera replication method: http://galeracluster.com/products/technology/ ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation
On Aug 10, 2014, at 11:17 AM, Li Ma skywalker.n...@gmail.com wrote: How about Galera multi-master cluster? As Mike Bayer said, it is virtually synchronous by default. It is still possible that outdated rows are queried that make results not stable. not sure if I said that :). I know extremely little about galera. Let's move forward to synchronous replication, like Galera with causal-reads on. The dominant advantage is that it has consistent relational dataset support. The disadvantage are that it uses optimistic locking and its performance sucks (also said by Mike Bayer :-). For optimistic locking problem, I think it can be dealt with by retry-on-deadlock. It's not the topic here. I *really* don’t think I said that, because I like optimistic locking, and I’ve never used Galera ;). Where I am ignorant here is of what exactly occurs if you write some rows within a transaction with Galera, then do some reads in that same transaction. I’d totally guess that Galera would need to first have SELECTs come from a slave node, then the moment it sees any kind of DML / writing, it transparently switches the rest of the transaction over to a writer node. No idea, but it has to be something like that? So, the transparent read/write separation is dependent on such an environment. SQLalchemy tutorial provides code sample for it [1]. Besides, Mike Bayer also provides a blog post for it [2]. So this thing with the “django-style routers”, the way that example is, it actually would work poorly with a Session that is not in “autocommit” mode, assuming you’re working with regular old databases that are doing some simple behind-the-scenes replication. Because again, if you do a flush, those rows go to the master, if the transaction is still open, then reading from the slaves you won’t see the rows you just inserted.So in reality, that example is kind of crappy, if you’re in a transaction (which we are) you’d really need to be doing session.using_bind(“master”) all over the place, and that is already way too verbose and hardcoded. I’m wondering why I didn’t make a huge note of that in the post. The point of that article was more to show that hey, you *can* control it at this level if you want to but you need to know what you’re doing. Just to put it out there, this is what I think good high/level master/slave separation in the app level (reiterating: *if we want it in the app level at all*) should approximately look like: @transaction.writer def read_and_write_something(arg1, arg2, …): # … @transaction.reader def only_read_something(arg1, arg2, …): # … that way there is no awareness of master/slave anything, the underlying system can decide what “reader” and “writer” means. Do in-app switching between two databases, send out some magic signals to some commercial clustering service, have the “readers” work in “autocommit” mode, or do nothing, whatever. The code doesn’t decide this imperatively.But it isn’t 100% “transparent”, this small amount of declaration per-method is needed. What I did is to re-implement it in OpenStack DB API modules in my development environment, using Galera cluster(causal-reads on). It has been running perfectly for more than a week. The routing session manager works well while maintaining data consistency. OK so Galera would perhaps have some way to make this happen, and that’s great. My understanding is that people are running Openstack already with Galera, that’s why we’re hitting issues with some of those SELECT..FOR UPDATEs that are being replaced with optimistic approaches as you mention. But beyond that this isn’t any kind of “change” to oslo.db or anything else. Run Openstack with whatever database backend you want, ideally (that is my primary agenda, sorry MySQL vendors!). Finally, I think if we can integrate it into oslo.db, it is a perfect plus for those who would like to deploy Galera (or other similar technology) as DB backend. this (the word “integrate”, and what does that mean) is really the only thing making me nervous. If the integration here is the django blog post I have, it’s not going to work with transactions. Either the system is magical enough that a single transaction can read/write from both sources midway and there is no “integration” needed, or the transaction has to be declared up front as reader or writer. Or you don’t use transactions except for writers, which is essentially the same as “declaration up front”. [1] http://docs.sqlalchemy.org/en/rel_0_9/orm/session.html#custom-vertical-partitioning [2] http://techspot.zzzeek.org/2012/01/11/django-style-database-routers-in-sqlalchemy/ [3] Galera replication method: http://galeracluster.com/products/technology/ ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org
Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation
not sure if I said that :). I know extremely little about galera. Hi Mike Bayer, I'm so sorry I mistake you from Mike Wilson in the last post. :-) Also, say sorry to Mike Wilson. I’d totally guess that Galera would need to first have SELECTs come from a slave node, then the moment it sees any kind of DML / writing, it transparently switches the rest of the transaction over to a writer node. You are totally right. @transaction.writer def read_and_write_something(arg1, arg2, …): # … @transaction.reader def only_read_something(arg1, arg2, …): # … The first approach that I had in mind is the decorator-based method to separates read/write ops like what you said. To some degree, it is almost the same app-level approach to the master/slave configuration, due to transparency to developers. However, as I stated before, the current approach is merely used in OpenStack. Decorator is more friendly than use_slave_flag or something like that. If ideally transparency cannot be achieved, to say the least, decorator-based app-level switching is a great improvement, compared with the current implementation. OK so Galera would perhaps have some way to make this happen, and that's great. If any Galera expert here, please correct me. At least in my experiment, transactions work in that way. this (the word “integrate”, and what does that mean) is really the only thing making me nervous. Mike, just feel free. What I'd like to do is to add a django-style routing method as a plus in oslo.db, like: [database] # Original master/slave configuration master_connection = slave_connection = # Only Support Synchronous Replication enable_auto_routing = True [db_cluster] master_connection = master_connection = ... slave_connection = slave_connection = ... HOWEVER, I think it needs more investigation, so this is why I'd like to put it in the mailing list in the early stage to raise some discussions in depth. I'm not a Galera expert. I really appreciate any challenges here. Thanks, Li Ma - Original Message - From: Mike Bayer mba...@redhat.com To: OpenStack Development Mailing List (not for usage questions) openstack-dev@lists.openstack.org Sent: 星期日, 2014年 8 月 10日 下午 11:57:47 Subject: Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation On Aug 10, 2014, at 11:17 AM, Li Ma skywalker.n...@gmail.com wrote: How about Galera multi-master cluster? As Mike Bayer said, it is virtually synchronous by default. It is still possible that outdated rows are queried that make results not stable. not sure if I said that :). I know extremely little about galera. Let's move forward to synchronous replication, like Galera with causal-reads on. The dominant advantage is that it has consistent relational dataset support. The disadvantage are that it uses optimistic locking and its performance sucks (also said by Mike Bayer :-). For optimistic locking problem, I think it can be dealt with by retry-on-deadlock. It's not the topic here. I *really* don’t think I said that, because I like optimistic locking, and I’ve never used Galera ;). Where I am ignorant here is of what exactly occurs if you write some rows within a transaction with Galera, then do some reads in that same transaction. I’d totally guess that Galera would need to first have SELECTs come from a slave node, then the moment it sees any kind of DML / writing, it transparently switches the rest of the transaction over to a writer node. No idea, but it has to be something like that? So, the transparent read/write separation is dependent on such an environment. SQLalchemy tutorial provides code sample for it [1]. Besides, Mike Bayer also provides a blog post for it [2]. So this thing with the “django-style routers”, the way that example is, it actually would work poorly with a Session that is not in “autocommit” mode, assuming you’re working with regular old databases that are doing some simple behind-the-scenes replication. Because again, if you do a flush, those rows go to the master, if the transaction is still open, then reading from the slaves you won’t see the rows you just inserted.So in reality, that example is kind of crappy, if you’re in a transaction (which we are) you’d really need to be doing session.using_bind(“master”) all over the place, and that is already way too verbose and hardcoded. I’m wondering why I didn’t make a huge note of that in the post. The point of that article was more to show that hey, you *can* control it at this level if you want to but you need to know what you’re doing. Just to put it out there, this is what I think good high/level master/slave separation in the app level (reiterating: *if we want it in the app level at all*) should approximately look like: @transaction.writer def read_and_write_something(arg1, arg2, …): # … @transaction.reader def only_read_something(arg1, arg2
Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation
Hi Li, How are you going to make this separation transparent? I mean, generally, in a function code, you can't know in advance if the transaction will be read-only or it will contain an INSERT/UPDATE/DELETE statement. On the other hand, as a developer, you could analyze the DB queries that can be possibly issued by this function and mark the function somehow, so that oslo.db would know for which database connection the transaction should be created, but this is essentially what slave_connection option is for and how it works now. Secondly, as you said, the key thing here is to separate reads and writes. In order to make reads fast/reduce the load on your 'writable' database, you'd move reads to asynchronous replicas. But you can't do this transparently either, as there is a lot of places in our code, in which we assume we are using the latest state of data, while asynchronous replicas might actually be a little bit out of date. So, in case of slave_connection, we use it only when it's ok for the code to work with outdated rows, i.e. *explicitly* modify the existing functions to work with slave_connection. Thanks, Roman On Fri, Aug 8, 2014 at 7:03 AM, Li Ma skywalker.n...@gmail.com wrote: Getting a massive amount of information from data storage to be displayed is where most of the activity happens in OpenStack. The two activities of reading data and writing (creating, updating and deleting) data are fundamentally different. The optimization for these two opposite database activities can be done by physically separating the databases that service these two different activities. All the writes go to database servers, which then replicates the written data to the database server(s) dedicated to servicing the reads. Currently, AFAIK, many OpenStack deployment in production try to take advantage of MySQL (includes Percona or MariaDB) multi-master Galera cluster. It is possible to design and implement a read/write separation schema for such a DB cluster. Actually, OpenStack has a method for read scalability via defining master_connection and slave_connection in configuration, but this method lacks of flexibility due to deciding master or slave in the logical context(code). It's not transparent for application developer. As a result, it is not widely used in all the OpenStack projects. So, I'd like to propose a transparent read/write separation method for oslo.db that every project may happily takes advantage of it without any code modification. Moreover, I'd like to put it in the mailing list in advance to make sure it is acceptable for oslo.db. I'd appreciate any comments. br. Li Ma ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation
On Aug 8, 2014, at 12:03 AM, Li Ma skywalker.n...@gmail.com wrote: So, I'd like to propose a transparent read/write separation method for oslo.db that every project may happily takes advantage of it without any code modification. A single transaction begins, which is to emit a series of INSERT, UPDATE, and SELECT statements. Are you proposing that this system in fact produce two separate transactions on two separate backends, and deliver the SELECT statements to the slave? That approach isn’t feasible - SELECTs are part of a “write” transaction just as much as the other statements are (as they can be SELECTing locally uncommitted data), as they deliver data which is part of the transactional context as well as intended for those DML statements. Otherwise, by what system could this read/write be “transparent”? reader/writer has to be at the transaction level, not the statement level, and without an up-front declaration as to whether a transaction is to be reader or writer, it’s not possible. ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [oslo.db]A proposal for DB read/write separation
Li Ma, This is interesting, In general I am in favor of expanding the scope of any read/write separation capabilities that we have. I'm not clear what exactly you are proposing, hopefully you can answer some of my questions inline. The thing I had thought of immediately was detection of whether an operation is read or write and integrating that into oslo.db or sqlalchemy. Mike Bayer has some thoughts on that[1] and there are other approaches around that can be copied/learned from. These sorts of things are clear to me and while moving towards more transparency for the developer, still require context. Please, share with us more details on your proposal. -Mike [1] http://www.percona.com/doc/percona-xtradb-cluster/5.5/wsrep-system-index.html [2] http://techspot.zzzeek.org/2012/01/11/django-style-database-routers-in-sqlalchemy/ On Thu, Aug 7, 2014 at 10:03 PM, Li Ma skywalker.n...@gmail.com wrote: Getting a massive amount of information from data storage to be displayed is where most of the activity happens in OpenStack. The two activities of reading data and writing (creating, updating and deleting) data are fundamentally different. The optimization for these two opposite database activities can be done by physically separating the databases that service these two different activities. All the writes go to database servers, which then replicates the written data to the database server(s) dedicated to servicing the reads. Currently, AFAIK, many OpenStack deployment in production try to take advantage of MySQL (includes Percona or MariaDB) multi-master Galera cluster. It is possible to design and implement a read/write separation schema for such a DB cluster. I just want to clarify here, are you suggesting that _all_ reads and _all_ writes would hit different databases? It would be interesting to see a relational schema design that would allow that to work. That seems like something that you wouldn't try in a relational database at all. Actually, OpenStack has a method for read scalability via defining master_connection and slave_connection in configuration, but this method lacks of flexibility due to deciding master or slave in the logical context(code). It's not transparent for application developer. As a result, it is not widely used in all the OpenStack projects. So, I'd like to propose a transparent read/write separation method for oslo.db that every project may happily takes advantage of it without any code modification. The problem with making it transparent to the developer is that, well, you can't unless your application is tolerant of old data in an asynchronous replication world. If you are in a fully synchronous world you could fully separate writes and reads, but what would be the point since your database performance is now trash anyway. Please note that although Galera is a considered a synchronous model it's not actually all the way there. You can break the certification of course, but there are also things that are done to keep the performance to an acceptable level. Take for example the wswrep_causal_reads configuration parameter[2]. Without this sucker being turned on you can't make read/write separation transparent to the developer. Turning it on causes a significant performance degradation unfortunately. I feel like this is a problem fundamental to a consistent relational dataset. If you are okay with eventual consistency it's okay, you can make things transparent to the developer. But by it's very nature relational datasets are well, relational, they need all the other pieces and those pieces need to be consistent. I guess what I am saying is that your proposal needs more details. Please respond with specifics and examples to move the discussion forward. Moreover, I'd like to put it in the mailing list in advance to make sure it is acceptable for oslo.db. I'd appreciate any comments. br. Li Ma ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
[openstack-dev] [oslo.db]A proposal for DB read/write separation
Getting a massive amount of information from data storage to be displayed is where most of the activity happens in OpenStack. The two activities of reading data and writing (creating, updating and deleting) data are fundamentally different. The optimization for these two opposite database activities can be done by physically separating the databases that service these two different activities. All the writes go to database servers, which then replicates the written data to the database server(s) dedicated to servicing the reads. Currently, AFAIK, many OpenStack deployment in production try to take advantage of MySQL (includes Percona or MariaDB) multi-master Galera cluster. It is possible to design and implement a read/write separation schema for such a DB cluster. Actually, OpenStack has a method for read scalability via defining master_connection and slave_connection in configuration, but this method lacks of flexibility due to deciding master or slave in the logical context(code). It's not transparent for application developer. As a result, it is not widely used in all the OpenStack projects. So, I'd like to propose a transparent read/write separation method for oslo.db that every project may happily takes advantage of it without any code modification. Moreover, I'd like to put it in the mailing list in advance to make sure it is acceptable for oslo.db. I'd appreciate any comments. br. Li Ma ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev