Good CTE, Bad CTE

203 points by radimm 49 days ago | 45 comments

vlaaad 48 days ago |

Use the term, never define the term, classic.

CTE stands for Common Table Expressions in SQL. They are temporary result sets defined within a single query using the WITH clause, acting like named subqueries to improve readability and structure.

radimm 48 days ago | |

OP here, damn - that's a very good point. Can't believe I missed it.

iainmerrick 48 days ago | | |

From the headline, I thought it might be about sports-related concussions!

I was morbidly curious what a "good CTE" could possibly be...

swasheck 48 days ago | | |

i appreciate the way you took the feedback. i saw the domain name and immediately knew the content and context. the article did not disappoint. i come from a heavy mssql background with some postgres sprinkled in, but my current company is migrating our mssql fleet to pg and it’s nice to have a technical foundation and article to be able to truly understand and pass the differences and similarities then how the two platforms handle workload. Traditionally it’s just been multiple sides, mercilessly criticizing each other for their deficiencies, but each platform has its own strengths and its own gaps. I’m excited to be a part of this migration, and I appreciate technical articles such as this that help me articulate the broader challenges to both our executive level levels, and our developers.

edit: syntax. voice to text was liberal with the comma abuse

xxs 48 days ago | | |

I read the article before the abbreviate definition inclusion as its very opening. I had never met the abbreviation before.

It'd be quite surprising the WITH statement in top a query to be the first feature to learn/use past basic SQL. Is it personal experience in some industry?

mcdonje 48 days ago | |

While generally a fair critique, the site does have "SQL" in its name.

tclancy 48 days ago | |

Agreed. I was relieved to see this wasn’t written by Cam Skatteboro.

lizknope 48 days ago | | |

Yeah, I thought I was on a sports site. Cam thinks CTE isn't real and is "all in your head." Technically correct about the all in your head part.

da_chicken 47 days ago | |

Eh, almost every link on the homepage has an initialism or acronym in the title, and roughly none of them are actually defining the term they're using. Indeed, not to point fingers, your own submissions make the same mistake.

Sure, yes, OP should (and now has) defined the term. But at the same time it's reasonable to expect that someone reading a blog post on BoringSQL.com would already know the term just as much as we could expect people interested in Clojure would know what a REPL is.

cpfohl 48 days ago | |

Your comment could have been more helpful without the first sentence. SAME content, same correction, less superiority: "CTE is an overloaded term, in this article ......"

This is a valuable comment, don't ruin it with sarcasm and rudeness.

NooneAtAll3 48 days ago | | |

> Your comment could have been more helpful without the first sentence.

no it wouldn't

the whole point is to critique the post

atombender 47 days ago |

I mostly use CTEs for organization these days, and in rare cases to express queries which cannot be written without them.

These days I often write queries like this (especially when doing exploratory ad hoc queries, but also in apps) even when it's not necessary to use a CTE:

    WITH
      a AS (
        SELECT ... FROM ... WHERE ... etc.
      )
    SELECT * FROM a

The first CTE query defines the input, and the main query just reads from it. Subsequent subqueries invoke steps on that input to group, filter, join, and so on.

This has a bunch of nice benefits. For example, it allows me to add steps incrementally, and to "comment out" a step I can simply change the next step's input to read from the preceding step. Each step can be read and understood in isolation.

I work a lot with Postgres, ClickHouse, and SQLite, and generally find that the database inlines and optimizes CTEs, and challenges mostly concern performance traps (like IN or EXISTS) that allly to non-CTE situations as well.

yen223 48 days ago |

I've always thought of CTEs as a code organisation tool, not an optimisation tool. The fact the some rdbms treats them as an optimisation fence was a bug, not a feature.

solumunus 48 days ago | |

Improved readability is definitely the primary benefit.

oveja 48 days ago |

To the author of the article. This was a really nice and educating read. You made me finally understand recursive CTEs, the org chart was a really good example. Thank you.

mcdonje 48 days ago |

There's some good stuff in here. I didn't know about the issues an aggregation in a CTE can cause and haven't used EXISTS much.

Regarding recursive CTEs, you might be interested in how DuckDb evolved them with USING KEY: https://duckdb.org/2025/05/23/using-key

scythmic_waves 48 days ago | |

That was also a great read, thanks!

bob1029 48 days ago |

> Recursive CTEs use an iterative working-table mechanism. Despite the name, they aren't truly recursive. PostgreSQL doesn't "call itself" by creating a nested stack of unfinished queries.

If you want something that is more like actual recursion (I.e., depth-first), Oracle has CONNECT BY which does not require the same kind of tracking. It also comes with extra features to help with cycle detection, stack depth reflection, etc.

If your problem is aligned with the DFS model, the oracle technique can run circles around recursive CTEs. Anything with a deep hierarchy and early termination conditions is a compelling candidate.

hans_castorp 48 days ago | |

> If you want something that is more like actual recursion (I.e., depth-first), Oracle has CONNECT BY which does not require the same kind of tracking. It also comes with extra features to help with cycle detection, stack depth reflection, etc.

All that is supported with CTEs as well. And both Postgres and Oracle support the SQL standard for these things.

You can't choose between breadth first/depth first using CONNECT BY in Oracle. Oracle's manual even states that CTE are more powerful than CONNECT BY

dspillett 48 days ago |

I wrangle databases by day, and do martial arts of an evening. Two arenas where CTEs can cause significant headaches!

uwemaurer 48 days ago |

Great article, I always like to structure my queries with CTEs and I was (wrongly) assuming it all gets inlined at the end. Sometimes it also gets complicated since these intermediate results can't be easily seen in a SQL editor. I was working on a UI to parse CTE queries and then execute them step by step to show the results of all the CTEs for easier understanding of the query (as part of this project https://github.com/sqg-dev/sqg/)

siddboots 48 days ago | |

I think your assumption about inlining is essentially correct. As far as I know postgres was the last major rdbms to have an optimiser fence around CTEs.

nraynaud 48 days ago | | |

I concur, “the Germans” have created an algorithm that completely “see through” subqueries/CTEs when planning a query. The way the query is written has no bearing on the execution.

qwertydog 48 days ago |

Great post - thanks. I think the columns in the index you suggested in the pre-pg12 section are in the wrong order (that index would get used)

radimm 48 days ago | |

Thanks - I will recheck later today.

jpalomaki 48 days ago |

Sometimes easy performance trick is to split the CTE to separate queries, put the results to unlogged temporary tables and add whatever indexes the next step needs.

Obviously makes only sense for stuff like analytical queries that are not running constantly.

taeric 48 days ago | |

Worth underlying the OLAP versus OLTP divide you are speaking to on the close, there.

andsbf 47 days ago | |

An issue that has arise for me in some situations is that for more expensive/reporting queries we point to a db replica, where temporary tables are not an option.

faangguyindia 47 days ago |

I consulted at some Indian startups. Such bad queries were written which were costing them so much compute

I just rewrote all queries with claude code and next day and compute decreased to 1/5th.

listenallyall 47 days ago | |

Kind of a weird flex - you're basically saying that you aren't any better than the Indian teams at writing efficient queries or optimizing existing ones, the only talent you've demonstrated is ability to ask Claude Code. It's like you are announcing your own obsolescence.

faangguyindia 47 days ago | | |

It's true, claude is already outperforming me

ctippett 48 days ago |

Not to sound like a broken record, but I need to echo what many other replies have already said: great article.

Embarassingly, despite thinking of myself as pretty knowledgeable with SQL, I had no idea you could nest DML statements inside a CTE. I always assumed/used DML statements as the final statement after a CTE was defined. I'm not sure if or when I might use this in the future, but it's neat to learn something new (and to be humbled at the same time).

intrasight 47 days ago |

I just wish that SQL Server had a materialize keyword like Postgres.

I'll write some nice clean CTEs and then have to refactor it to temp tables due to the lack of materialization.

jt-hill 48 days ago |

Wow it just keeps going. Extremely thorough.

BSTRhino 47 days ago |

Love CYEs, use them all the time, great article.

BSTRhino 47 days ago | |

Ah typo, sorry CTEs!