5 ms·

PostgreSQL, pg_shard, and what we learned from our failures

108 points by craigkerstiens 11y ago

rch 11y ago
> The problem was that we were trying to retrofit [the Foreign Data Wrapper] API for something that it was fundamentally not designed to do. > The PostgreSQL Extension APIs ... enable you to extend, cooperate with, or override any database submodule's behavior. We could therefore change our design to use the planner and executor hook APIs, and we found that things followed nicely from there.
- tinco 11y ago
  Unfortunately this comment is really a good summary of the article. I guess it makes marketing sense to cut up big blog posts into two, but as it is the article really is vapid, while it could easily have been the article that makes me consider CitusDB over MongoDB for a particular application.
  - rch 11y ago
    To be fair, the details of why statement one is true are still interesting though.
mercurial 11y ago
I think PG is a great database, but I don't see what makes it a good fit for a distributed system. You're going to either give up too much performance or a lot of the guarantees that make standard SQL databases interesting in the first place.
- ozgune 11y ago
  (Ozgun from Citus Data) It's Postgres' Extension APIs. They enable you to extend, cooperate with, or override any database submodule's behavior without forking the database: https://goo.gl/rr2EIm https://goo.gl/rr2EIm As an example, let's say you leveraged planner hooks to parallelize PostgreSQL queries. When a new feature such as JSONB gets checked into Postgres, it becomes immediately available for parallel querying. To my knowledge, no other database provides the means to extend the database to such an extent. These APIs became official in PG 9.0, and two example links from the PostgreSQL documentation are: http://www.postgresql.org/docs/9.4/static/extend.html http://www.postgresql.org/docs/9.4/static/extend.html http://www.postgresql.org/docs/9.4/static/spi.html http://www.postgresql.org/docs/9.4/static/spi.html
  - mercurial 11y ago
    This doesn't really answer my question. I'm not asking "how can you turn PG into a distributed database", I'm asking "what do I get with that that I don't get with, eg, Cassandra"?
    - tracker1 11y ago
      Structured Query Language ... that's the biggest thing you get. Also database resolved relational data. It also allows you to present said information from other stores through a well defined interface, such as via stored proceedures that can be locked down via login/identity/application and other features typically in SQL, while being able to act as a gateway for more optimized data stores for specific types of data. Not to mention that JSON data in pgsql has gotten pretty good... I think with plv8 it gets event better. Where I think pg falls is in terms of replica sets, hot failover and automated recovery options which many nosql servers give you in the box. With PG you pretty much have to pay for a decent HA solution. That said, I honestly really like document stores, and RethinkDB as an example offers some of the relational features of a typical SQL database while being mainly available as a document store.
      - threeseed 11y ago
        You can get SQL over Cassandra or almost everything other database (even MongoDB) via Spark. And CQLSH is SQL-like at least. And personally I have always thought putting business logic in stored procedures is a truly dumb idea. You forever lock yourself to that database and there is always a point that you can exceed its capabilities and need to implement functionality in the application layer anyway.
        qaqy 11y ago
        There is thins annoying thing called security with such things as trust boundaries etc. So if you are creating a mildly secure web app the only permission your web apps db user has is to execute a subset of stored procedures and that is it.
jabbernotty 11y ago
I find the proposed solution for 'SELECT COUNT(customer_id) FROM orders' confusing. They suggest placing the 3 billion order rows into three sets of 1 billion, to be distributed among three machines. But summing the number of distinct customer IDs in each bucket doesn't give you the number of distinct customer IDs in the orders table because an ID could be in more than one bucket. What am I missing?
- lobster_johnson 11y ago
  Indeed, that's confusing. You don't want to push down the count; the plan must execute "select distinct customer_id" separately on each shard, then merge the results and count. Not sure if that's what they really meant — since their business is sharding, this shouldn't be new to them.
- ozgune 11y ago
  This isn't clear in the blog post; I'll try to clarify. The buckets correspond to hash token ranges. For example, the database takes one row from orders_2013, hashes the customer_id value, and gets a hashed value of 1.5B. It then puts it into the second bucket orders_2013_[1B-2B[. After the distributed shuffle, you end up with orders_cust_id[1B-2B[ on machine 2. So all orders that belong to a particular / disjoint set of customers live on one machine. You can then push down the count(distinct customer_id) to each node, and then add up the results. We use hash token ranges in pg_shard. I'll include this clarification to the blog post.
dharbin 11y ago
Should Citus reevaluate cstore_fdw in light of this extensions API? I wonder if a foreign data wrapper is similarly ill-fitted for this application.
- mslot 11y ago
  For basic reads and writes the FDW APIs are quite suitable and they work well for cstore_fdw. However, to get the most out of columnar storage you also need vectorized execution that can calculate aggregates on blocks of values. The FDW APIs are not well-suited for this purpose, since they pull one row of data at a time. We've done some experiments in which we used the executor hooks to implement a vectorized executor for cstore_fdw [1] and it can give dramatic performance gains. However, the SQL support of this solution is limited as the APIs force you to build a new executor from scratch. With sharding, the distributed executor can work on the SQL level and translate one SQL query to many SQL queries. In that case your distributed executor can focus on issues such as partitioning and merging data, while the low-level work is still being done by the regular PostgreSQL executor, so you don't have to implement it from scratch. [1] https://github.com/citusdata/postgres_vectorization_test https://github.com/citusdata/postgres_vectorization_test

108 points by craigkerstiens 11y ago

rch 11y ago

> The problem was that we were trying to retrofit [the Foreign Data Wrapper] API for something that it was fundamentally not designed to do. > The PostgreSQL Extension APIs ... enable you to extend, cooperate with, or override any database submodule's behavior. We could therefore change our design to use the planner and executor hook APIs, and we found that things followed nicely from there.

tinco 11y ago
Unfortunately this comment is really a good summary of the article. I guess it makes marketing sense to cut up big blog posts into two, but as it is the article really is vapid, while it could easily have been the article that makes me consider CitusDB over MongoDB for a particular application.
- rch 11y ago
  To be fair, the details of why statement one is true are still interesting though.

mercurial 11y ago

I think PG is a great database, but I don't see what makes it a good fit for a distributed system. You're going to either give up too much performance or a lot of the guarantees that make standard SQL databases interesting in the first place.

ozgune 11y ago
(Ozgun from Citus Data) It's Postgres' Extension APIs. They enable you to extend, cooperate with, or override any database submodule's behavior without forking the database: https://goo.gl/rr2EIm https://goo.gl/rr2EIm As an example, let's say you leveraged planner hooks to parallelize PostgreSQL queries. When a new feature such as JSONB gets checked into Postgres, it becomes immediately available for parallel querying. To my knowledge, no other database provides the means to extend the database to such an extent. These APIs became official in PG 9.0, and two example links from the PostgreSQL documentation are: http://www.postgresql.org/docs/9.4/static/extend.html http://www.postgresql.org/docs/9.4/static/extend.html http://www.postgresql.org/docs/9.4/static/spi.html http://www.postgresql.org/docs/9.4/static/spi.html
- mercurial 11y ago
  This doesn't really answer my question. I'm not asking "how can you turn PG into a distributed database", I'm asking "what do I get with that that I don't get with, eg, Cassandra"?
  - tracker1 11y ago
    Structured Query Language ... that's the biggest thing you get. Also database resolved relational data. It also allows you to present said information from other stores through a well defined interface, such as via stored proceedures that can be locked down via login/identity/application and other features typically in SQL, while being able to act as a gateway for more optimized data stores for specific types of data. Not to mention that JSON data in pgsql has gotten pretty good... I think with plv8 it gets event better. Where I think pg falls is in terms of replica sets, hot failover and automated recovery options which many nosql servers give you in the box. With PG you pretty much have to pay for a decent HA solution. That said, I honestly really like document stores, and RethinkDB as an example offers some of the relational features of a typical SQL database while being mainly available as a document store.
    - threeseed 11y ago
      You can get SQL over Cassandra or almost everything other database (even MongoDB) via Spark. And CQLSH is SQL-like at least. And personally I have always thought putting business logic in stored procedures is a truly dumb idea. You forever lock yourself to that database and there is always a point that you can exceed its capabilities and need to implement functionality in the application layer anyway.
      - qaqy 11y ago
        There is thins annoying thing called security with such things as trust boundaries etc. So if you are creating a mildly secure web app the only permission your web apps db user has is to execute a subset of stored procedures and that is it.

jabbernotty 11y ago

I find the proposed solution for 'SELECT COUNT(customer_id) FROM orders' confusing. They suggest placing the 3 billion order rows into three sets of 1 billion, to be distributed among three machines. But summing the number of distinct customer IDs in each bucket doesn't give you the number of distinct customer IDs in the orders table because an ID could be in more than one bucket. What am I missing?

lobster_johnson 11y ago
Indeed, that's confusing. You don't want to push down the count; the plan must execute "select distinct customer_id" separately on each shard, then merge the results and count. Not sure if that's what they really meant — since their business is sharding, this shouldn't be new to them.
ozgune 11y ago
This isn't clear in the blog post; I'll try to clarify. The buckets correspond to hash token ranges. For example, the database takes one row from orders_2013, hashes the customer_id value, and gets a hashed value of 1.5B. It then puts it into the second bucket orders_2013_[1B-2B[. After the distributed shuffle, you end up with orders_cust_id[1B-2B[ on machine 2. So all orders that belong to a particular / disjoint set of customers live on one machine. You can then push down the count(distinct customer_id) to each node, and then add up the results. We use hash token ranges in pg_shard. I'll include this clarification to the blog post.

dharbin 11y ago

Should Citus reevaluate cstore_fdw in light of this extensions API? I wonder if a foreign data wrapper is similarly ill-fitted for this application.

mslot 11y ago
For basic reads and writes the FDW APIs are quite suitable and they work well for cstore_fdw. However, to get the most out of columnar storage you also need vectorized execution that can calculate aggregates on blocks of values. The FDW APIs are not well-suited for this purpose, since they pull one row of data at a time. We've done some experiments in which we used the executor hooks to implement a vectorized executor for cstore_fdw [1] and it can give dramatic performance gains. However, the SQL support of this solution is limited as the APIs force you to build a new executor from scratch. With sharding, the distributed executor can work on the SQL level and translate one SQL query to many SQL queries. In that case your distributed executor can focus on issues such as partitioning and merging data, while the low-level work is still being done by the regular PostgreSQL executor, so you don't have to implement it from scratch. [1] https://github.com/citusdata/postgres_vectorization_test https://github.com/citusdata/postgres_vectorization_test