An Overview of Sharding in PostgreSQL and How it Relates to MongoDB’s

PostgreSQL LogoA couple of weeks ago I presented at Percona University São Paulo about the new features in PostgreSQL that allow the deployment of simple shards. I’ve tried to summarize the main points in this post, as well as providing an introductory overview of sharding itself. Please note I haven’t included any third-party extensions that provide sharding for PostgreSQL in my discussion below.

Partitioning in PostgreSQL

In a nutshell, until not long ago there wasn’t a dedicated, native feature in PostgreSQL for table partitioning. Not that that prevented people from doing it anyway: the PostgreSQL community is very creative. There’s a table inheritance feature in PostgreSQL that allows the creation of child tables with the same structure as a parent table. That, combined with the employment of proper constraints in each child table along with the right set of triggers in the parent table, has provided practical “table partitioning” in PostgreSQL for years (and still works). Here’s an example:

Using table inheritance

Figure 1a. Main (or parent) table

Figure 1b. Child tables inherit the structure of the parent table and are limited by constraints

Figure 1c. A function that controls in which child table a new entry should be added according to the timestamp field

Figure 1d. A trigger is added to the parent table that calls the function above when an INSERT is performed

The biggest drawbacks for such an implementation was related to the amount of manual work needed to maintain such an environment (even though a certain level of automation could be achieved through the use of 3rd party extensions such as pg_partman) and the lack of optimization/support for “distributed” queries. The PostgreSQL optimizer wasn’t advanced enough to have a good understanding of partitions at the time, though there were workarounds that could be used such as employing constraint exclusion.

Declarative partitioning

About 1.5 year ago, PostgreSQL 10 was released with a bunch of new features, among them native support for table partitioning through the new declarative partitioning feature. Here’s how we could partition the same temperature table using this new method:

Figure 2a. Main table structure for a partitioned table

Figure 2b. Tables defined as partitions of the main table; with declarative partitioning, there was no need for triggers anymore.

It still missed the greater optimization and flexibility needed to consider it a complete partitioning solution. It wasn’t possible, for example, to perform an UPDATE that would result in moving a row from one partition to a different one, but the foundation had been laid. Fast forward another year and PostgreSQL 11 builds on top of this, delivering additional features like:

  • the possibility to define a default partition, to which any entry that wouldn’t fit a corresponding partition would be added to.
  • having indexes added to the main table “replicated” to the underlying partitions, which improved declarative partitioning usability.
  • support for Foreign Keys

These are just a few of the features that led to a more mature partitioning solution.

Sharding in PostgreSQL

By now you might be reasonably questioning my premise, and that partitioning is not sharding, at least not in the sense and context you would have expected this post to cover. In fact, PostgreSQL has implemented sharding on top of partitioning by allowing any given partition of a partitioned table to be hosted by a remote server. The basis for this is in PostgreSQL’s Foreign Data Wrapper (FDW) support, which has been a part of the core of PostgreSQL for a long time. While technically possible to implement, we just couldn’t make practical use of it for sharding using the table inheritance + triggers approach. Declarative partitioning allowed for much better integration of these pieces making sharding – partitioned tables hosted by remote servers – more of a reality in PostgreSQL.

Figure 3a. On the remote server we create a “partition” – nothing but a simple table

Figure 3b. On the local server the preparatory steps involve loading the postgres_fdw extension, allowing our local application user to use that extension, creating an entry to access the remote server, and finally mapping that user with a user in the remote server (fdw_user) that has local access to the table we’ll use as a remote partition.

Figure 3c. Now it’s simply a matter of creating a proper partition of our main table in the local server that will be linked to the table of the same name in the remote server

You can read more about postgres_fdw in Foreign Data Wrappers in PostgreSQL and a closer look at postgres_fdw.

When does it make sense to partition a table?

There are a several principle reasons to partition a table:

  1. When a table grows so big that searching it becomes impractical even with the help of indexes (which will invariably become too big as well).
  2. When data management is such that the target data is often the most recently added and/or older data is constantly being purged/archived, or even not being searched anymore (at least not as often).
  3. If you are loading data from different sources and maintaining it as a data warehousing for reporting and analytics.
  4. For a less expensive archiving or purging of massive data that avoids exclusive locks on the entire table.

When should we resort to sharding?

Here are a couple of classic cases:

  1. To scale out (horizontally), when even after partitioning a table the amount of data is too great or too complex to be processed by a single server.
  2. Use cases where the data in a big table can be divided into two or more segments that would benefit the majority of the search patterns. A common example used to describe a scenario like this is that of a company whose customers are evenly spread across the United States and searches to a target table involves the customer ZIP code. A shard then could be used to host entries of customers located on the East coast and another for customers on the West coast.

Note though this is by no means an extensive list.

How should we shard the data?

With sharding (in this context) being “distributed” partitioning, the essence for a successful (performant) sharded environment lies in choosing the right shard key – and by “right” I mean one that will distribute your data across the shards in a way that will benefit most of your queries. In the example above, using the customer ZIP code as shard key makes sense if an application will more often be issuing queries that will hit one shard (East) or the other (West). However, if most queries would filter by, say, birth date, then all queries would need to be run through all shards to recover the full result set. This could easily backfire on performance with the shard approach, by not selecting the right shard key or simply by having such a heterogeneous workload that no shard key would be able to satisfy it.

It only ever makes sense to shard if the nature of the queries involving the target table(s) is such that distributed processing will be the norm and constitute an advantage far greater than any overhead caused by a minority of queries that rely on JOINs involving multiple shards. Due to the distributed nature of sharding such queries will necessarily perform worse if compared to having them all hosted on the same server.

Why not simply rely on replication or clustering?

Sharding should be considered in those situations where you can’t efficiently break down a big table through data normalization or use an alternative approach and maintaining it on a single server is too demanding. The table is then partitioned and the partitions distributed across different servers to spread the load across many servers. It doesn’t need to be one partition per shard, often a single shard will host a number of partitions.

Note how sharding differs from traditional “share all” database replication and clustering environments: you may use, for instance, a dedicated PostgreSQL server to host a single partition from a single table and nothing else. However, these data scaling technologies may well complement each other: a PostgreSQL database may host a shard with part of a big table as well as replicate smaller tables that are often used for some sort of consultation (read-only), such as a price list, through logical replication.

How does sharding in PostgreSQL relates to sharding in MongoDB®?

MongoDB® tackles the matter of managing big collections straight through sharding: there is no concept of local partitioning of collections in MongoDB. In fact, the whole MongoDB scaling strategy is based on sharding, which takes a central place in the database architecture. As such, the sharding process has been made as transparent to the application as possible: all a DBA has to do is to define the shard key.

Instead of connecting to a reference database server the application will connect to an auxiliary router server named mongos which will process the queries and request the necessary information to the respective shard. It knows which shard contains what because they maintain a copy of the metadata that maps chunks of data to shards, which they get from a config server, another important and independent component of a MongoDB sharded cluster. Together, they also play a role in maintaining good data distribution across the shards, actively splitting and migrating chunks of data between servers as needed.

In PostgreSQL the application will connect and query the main database server. There isn’t an intermediary router such as the mongos but PostgreSQL’s query planner will process the query and create an execution plan. When data requested from a partitioned table is found on a remote server PostgreSQL will request the data from it, as shown in the EXPLAIN output below:

Figure 4: excerpt of an EXPLAIN plan that involves processing a query in a remote server

Note in the above query the mention “Remote SQL”. A lot of optimizations have been made in the execution of remote queries in PostgreSQL 10 and 11, which contributed to mature and improve the sharding solution. Among them is support for having grouping and aggregation operations executed on the remote server itself (“push down”) rather than recovering all rows and having them processed locally.

What is missing in PostgreSQL implementation?

There is, however, still room for improvement. In terms of remote execution, reports from the community indicate not all queries are performing as they should. For example, in some cases the PostgreSQL planner is not performing a full push-down, resulting in shards transferring more data than required. Parallel scheduling of queries that touch multiple shards is not yet implemented: for now, the execution is taking place sequentially, one shard at a time, which takes longer to complete. When it comes to the maintenance of partitioned and sharded environments, changes in the structure of partitions are still complicated and not very practical. For example, when you add a new partition to a partitioned table with an appointed default partition you may need to detach the default partition first if it contains rows that would now fit in the new partition, manually move those to the new partition, and finally re-attach the default partition back in place.

But that is all part of a maturing technology. We’re looking forward to PostgreSQL 12 and what it will bring in the partitioning and sharding fronts.


Image based on photos by Leonardo Quatrocchi from Pexels

 

Share this post

Comments (2)

  • Kay Agahd Reply

    Well written and very interesting, thank you! However, you write:
    “It only ever makes sense to shard if the nature of the queries involving the target table(s) is such that distributed processing will be the norm […] Due to the distributed nature of sharding such queries will necessarily perform worse if compared to having them all hosted on the same server.”
    While I fully understand your point, I wonder why it shouldn’t be beneficial to have less data on each shard. So even if the query hits every shard, each shard has to work through fewer data (for 10 shards only one-tenth). Of course, to be beneficial, it requires that the query is sent to all shards in parallel, which is not yet implemented in PostgreSQL as you wrote at the very end of your article – but I think it will be implemented in a near future and already is the case for mongoDB since its early days.

    May 27, 2019 at 10:13 am
  • Anthony Reply

    Do you known the extension Citus ? https://www.citusdata.com/

    May 29, 2019 at 5:18 am

Leave a Reply