Postgres Insider Terminology
Last week Craig Kerstiens published a great introduction to Postgres terminology, covering some of the basics you might run into when just getting started.
The funny thing about jargon is how quickly we get used to it, and forget we are even using it. It becomes part of the secret handshake, the way we signal to other members of our tribe that we're part of the group.
When I first started going to Postgres conferences and listening to talks by Postgres core developers I suddenly found myself at sea. What were these strange words and phrases they were using?
It turns out that a lot of them are taken from the Postgres code base, which in turn naturally uses them because they are part of Codd's relational model, the theoretical information architecture that underpins all relational databases.
Here's an unofficial, incomplete glossary of weird terms that come out of developers mouths when talking about Postgres.
Can developers ever stop saying "tuple"? I cannot! It's a foundational building block, so it comes up a lot.
A "tuple" is what most people understand as a "row" in a "table". It has a number of attributes, each of which has a name, a value, and a data type.
Frequently you will also hear someone say "tuple set". A set of tuples is a set of rows, which is to say, a table! Or, at least, a "table-like thing". As we'll see later, there are also other special words for table-like things.
Record and Row
"Record" and "row" are another way of saying "tuple", but unlike "tuple", "record" and "row" are actually used in SQL when working with Postgres.
The "record" keyword is most frequently used as a return type in server side functions that return composite types.
CREATE FUNCTION onerow(pk integer) RETURNS record AS $$ SELECT * FROM mytable WHERE mytable.pk = pk $$ LANGUAGE 'sql';
The "row" keyword is used to construct composite (tuple valued) types within SQL.
INSERT INTO on_hand VALUES (ROW('fuzzy dice', 42, 1.99), 1000);
Arrays are a list of values, all of the same type. Here is an array of integers.
Arrays can also be multi-dimensional.
Most amazingly, arrays are not restricted to just the built-in Postgres types. You can make an array of any type at all!
CREATE TYPE person AS (name text, age integer); SELECT ('Peter',45)::person; SELECT ARRAY[('Peter',45),('Paul',35)]::person;
From Codd's original paper, a "relation" is a collection of tuples. A single tuple is a "unary relation", two tuples are a "binary relation", N tuples are an "nary" relation.
But we usually call relations "tables".
Every table is a relation. But not every relation is a table!
Views are also relations.
So, really, are the results of queries. A query result is a set of tuples, it is a relation! Even the intermediate steps in a query can be thought of as a relation.
The result of joining two tables using a join condition? Also a relation!
You can get a feel for how pervasive the idea of a "relation" is in Postgres by
pg_class system table like this:
SELECT relname, relkind FROM pg_class
This lists all the things Postgres considers to be "relations". You will see
relkind of "r") in the list. You will also see indexes, views,
sequences, materialized views, and foreign tables.
Target List and Restrictions
As you get deeper into the Postgres core, you will see things named that you
didn't know had a name. What is the name of the part of a SQL query after the
SELECT but before the
FROM? What is the name of the stuff after the
SELECT <this is the target list> FROM mytable WHERE <these are the restrictions>
The first part is the "target list". The second part are the "restrictions".
I have to include "schema" here because on the one hand it is very easy, Craig
already covered it in his
blog post: a
schema is what you get when you
CREATE SCHEMA! It's like a folder in your
But actually! Internally you will find that Postgres refers to "schemas" as
"namespaces". You can see a list of all the schemas in your database by running
a query on ...
SELECT * FROM pg_namespace
And that's not all. When I was first introduced to "schema" I was very confused, because to my mind a "schema" was a "collection of table definitions in a database". So an "application schema" might be all the tables in a database used to support a particular software application.
So, a "schema" is (to SQL) a folder where you can put database objects (like tables, functions, etc) and might (in Postgres) also be called a "namespace", and is sometimes (to an app developer) the definition of one or more tables.
Hang around Postgres developers long enough, they'll inevitably start talking about "pages".
Fortunately database end users can pretty safely ignore "pages", they are an implementation detail, and are never visible to users of SQL.
Database administrators should probably know what a "page" is, because some of the tuning parameters make references to pages.
Down at the level of the disk, database tables reside in files, one file per table (actually if the tables get large enough, they get multiple files, but let's ignore that for now).
If every tuple in the table were packed into a single file as tight as possible, how would the database update the data as it changed?
If it had to change someone's name from "Paul" to "Paul the Great", for example? It would have to expand the table, and move everything after the particular name entry down enough to make space for the new, larger entry. Not efficient!
Enter "pages". Rather than one tightly stuffed file of data, the database internally divides the table file into regularly spaced "pages", and only partially fills each page with tuples (there's another piece of jargon for this, the "fill factor"). Now, when the database needs to update a tuple, to perhaps add more data, only the page of interest needs to be re-written, and the rest of the file can remain untouched.
The "page" is in many ways the fundamental unit of the underlying database engine. You don't see it when writing SQL, but things like the "page cache" (a piece of RAM where frequently read pages are placed for high-speed access) and the "random page cost" (a tuning parameter that expresses how expensive random access within files is) testify to the centrality of the "page" in the system design of the database. PostgreSQL pages are fixed at exactly 8192 bytes. That is also the default size of memory “blocks” in areas like the WAL. Some block sizes can be increased; pages, never.
"Are those objects toasted?"
Toasted? What are you talking about? This is a database, not a bakery!
The Postgres documentation helpfully explains that "TOAST" is actually an acronym, for "The Oversized-Attribute Storage Technique".
Remember how "pages" are the building block of storage in Postgres? Well, they are a fixed size (8kb), so what happens if you want to store a record that is bigger than 8kb? Now we are in the realm of "oversized attributes".
Like "pages", "TOAST" is something that just magically works under the covers and is invisible to end users. When an incoming attribute is too large to fit into a page, Postgres slices it into smaller pieces that do fit inside a page, and puts those pieces into a special "TOAST table" associated with the main table.
-- Find the toast table associated with a user visible table SELECT a.relname, b.relname AS toast_relname FROM pg_class a JOIN pg_class b ON a.reltoastrelid = b.oid WHERE a.relname = 'mytablename';
Then, when you go to retrieve that tuple, the database goes to the TOAST table, gets the pieces and glues them back together again before returning them to you.
This tuple re-construction time is the main point of visibility of the TOAST system: retrieving a batch of TOASTed tuples will take longer than retrieving a similar batch of smaller tuples! Because (a) there is just more data in the TOASTed records and (b) gluing together the TOASTed pieces takes a non-zero amount of computational time.
This is just a few of the words that get thrown around when talking about Postgres and how it works, but I bet there are enough others that we will have an expanded glossary to publish soon!
November 3, 2022 •More by this author