Comment by brightball
7 years ago
I don't disagree that SQL can be improved. It's one of the biggest reasons I use Postgres in the first place because there are so many improvements available on top of SQL.
All that said...SQL is pretty darn effective. As a language, it's the true backbone of the internet today. It's readable, explicit, fairly concise and naturally translates to how data should be broken down for efficient storage...or make some trade-offs to allow for more efficient retrieval.
There are differences with different vendor implementations...but that's what different vendors are for - to find things the other guys are doing wrong and improving on them to build a better product.
I wish the folks luck in their work to improve things, but the language I've been able to rely on consistently over the last 17 years or so has been SQL...and I've worked with a lot of languages. SQL is the one that gets the most done reliably and lets me down the least often.
Afaict, the state of SQL as a grammar with tooling is kind of pathetic.
As a standardized language, it doesn’t really exist; everyone implements numerous extensions, and almost no one is fully ansi compliant
Almost all formatters attempt to be generic (believing standardization exists), and fail to support the full grammar for any dialect.
Across the board, all parsers have pathetic error message support (error on line 3, which is actually just the start of the statement).
The schema offers type constraints, but querying/ide’s extract no value from that (that is, types are statically specified/constrained, but query editors all pretend its fully dynamic)
Theres a lot of awkward nonsense, like where clauses are parsed before the select in most parsing engines, causing alias usage to fail without wrapping in a subselect/with clause
The grammars themselves are an inconsistent, ad-hoc mess
The grammar is also unnecessarily context-dependent (eg from must follow select, and where after that), making programmatic composition unnecessarily difficult
I don’t know how much of the tooling issue is a result of SQL as a language versus the history itself, but I can at least confirm that trying to parse multiple dialects is absolute hell, which would at least explain the sorry state of affairs for eg formatters.
But the majority of its expressive power derives from the relational algebra, and has nothing to do with the SQL grammar, and thats the majority of its value. It seems obvious to me that at the very least the compositional issue of SQL, and its self-inconsistent grammar, should be vulnerable to near-lossless improvement without too much struggle, though I can’t say what the alternative would actually look like.
But it seems like its riddled with a lot of unnecessary flaws
Actually standards compliance is really good these days. There are a lot of custom functions that are hard to do without in some reporting applications but behavior of SQL clauses is consistent across every engine.
JetBrains tooling does use schema metadata for it's autocompletion. I'm not sure though if it's anywhere close to full-on Haskell autocompletion in the Atom editor (very fiddly and prone to break on minor version changes, I must say).
Offtopic, but I am genuinely curious now about whether Atom's Haskell autocompletion is significantly better than in other editors like VSCode, or (n)vim with things like YouCompleteMe/NeoComplete/Deocomplete etc. Have you used other editors? What is your opinion?
The SQL spec is actually hidden away behind some book or something you have to purchase.
I was experimenting around with creating an Entity Framework equivalent in Typescript and really wanted to create a SQL AST for use under the hood(optimizing queries, SQL push down, etc). Ended up using the PostgreSQL types and a Ruby plugin that binds some PostgreSQL libs to work on my POC. Crazy town.
Big boys SQL is relatively close to the standard.
I’ve been writing some pretty ambitious Hive queries at work lately.
As I learn more about SQL and pull off more complex queries, my respect for it deepens. To have such power and support so many use cases with so few constructs is really an engineering feat. It’s timeless for a reason.
Some of my relatively common access patterns are awkward to express, but they can still be expressed in a few lines + a CTE or two, which is really impressive for a language so small.
This is not to say we can’t do better. But SQL has achieved a deep resonance with its problem space that most tools don’t even come close to. The brightest minds and most effective tooling shops in our field would be lucky merely to do as well.
This shows that humans can learn languages and get motivated by mastering them. This does not tell anything about the consistency, composability or orthogonality if SQL. Those qualities affecting the newcomers effort to learn it.
I think you mean the relational model is timeless. SQL is simply just not a great language, for an otherwise great idea.
Blindly asserting that SQL is bad without providing any argument or proof does not contradict OP's opinion on SQL.
26 replies →
Hive queries are written in HiveQL, not SQL. I used to write a lot of Hive and Impala queries, and going back to plain SQL is disappointing.
While based on SQL, HiveQL does not strictly follow the full SQL-92 standard, just like all the other SQL dialects out there. hiveQL is SQL.
2 replies →
It supports so many use cases primarily because of the small number of fundamental "constructs".
I mean, if I gave you protons, electrons and neutrons, you could build the universe out of them!
That doesn't make sense, if I give you more things (I'm avoiding saying 'atom', but you know what I mean) then it's not the case that you can suddenly do less.
Expressiveness isn't inversely correlated to number of constructs.
8 replies →
Depends on whether photons exist or not.
Radiation makes the universe a lot more interesting.
In your analogy, SQL is the periodic table.
Having used splunk's query language... I'd hope that in general vendors start with an assumption of SQL support and maybe try tweaking it at the edges rather than burning the house down - SQL has survived this long because it's extremely expressive and any replacement for it is going to need to match that expressiveness.
All that said I think it was originally structured to partially be a human readable language and it fails pretty hard at that - that's a facet I'm sure smart people could revise to make more natural.
To be fair, Splunk excels at dealing with unstructured logs, which is in many ways a harder problem. Obviously in a perfect world, all our data would be structured, which would make querying much simpler.
I used to work for Splunk. Querying Splunk with SQL is completely plausible, and something that Splunk has made a number of attempts at over the years.
The problem isn't SQL. It's that Splunk's query engine is tied up internally with a "grammar" that is a direct port of a shell pipeline into C++ with no intermediate representation or anything a compiler guy would recognize as a grammar. There was no design, no mathematical underpinning to it.
Splunk's unstructured log capabilities are really domain knowledge about making them semistructured as fast as possible: token indexing, a lot of effort on recognizing character encodings and timestamps intelligently, looking for key=value pairs, and letting people write regexes to extract fields themselves. The query language isn't somehow designed for a different data model.
In EWD1123, Dijkstra showed that the relational calculus and the regularity calculus (which governs regexes) are basically the same thing. My takeaway from that is that the relational model can be reinterpreted as a model over anything you want to match and manipulate with regexes by just changing the field selectors.
Typically the earlier part of your pipeline turns the logs into data records via regex captures. So you could have had a pipeline of what are essentially shell commands that produce a table that SQL operates on.
I don't buy that, splunk has a clear concept of columns and rows and could have chosen to expose support for a SQL like grammar for assembling them - they instead chose a less structured format that makes data assembly quite difficult.
More legitimately splunk may simply be unable to deliver performant data expression if a user is typing a query complex enough to justify SQL.
> I use Postgres […] SQL is pretty darn effective.
FWIW postgres used to have its own query language derived from QUEL[0] rather than SQL.
And findings that SQL is kinda shit are not exactly recent, e.g. C.J. Date's "A Critique of the SQL Language" (1983) lists the following sections
* lack of orthogonality: expressions
* lack of orthogonality: builtin functions
* lack of orthogonality: miscellaneous items formal definition
* mismatch with host languages
* missing function
* mistakes
* aspects of the relational model not supported
The conclusion was, obviously, prescient:
> if SQL is adopted on a wide scale in its present fortm~ then we will to some degree have missed the relational boat~ or at least failed to capitalize to the fullest possible extent on the potential of the relational model. That would be a pity, because we had an opportunity to do it right, and with a little effort we could have done so. The question is whether it is now too late. I sincerely hope not.
SQL succeeded not because it's "pretty darn effective" but because IBM decided on it (at a time where it drove technology) and Oracle are great at sales and marketing (whereas Ingres definitely wasn't).
[0] https://en.wikipedia.org/wiki/QUEL_query_languages
> I use Postgres in the first place because there are so many improvements available on top of SQL.
Most of Postgres is standard SQL. It's just that most non-Postgres databases do not implement standard SQL very well.
The various extensions / plugins that allow for custom data types, indexes, use of multiple programming languages to write functions, ability to use a foreign data wrapper to connect to Redis and build a VIEW out of the result or push data out to Redis/Memcached with a database function, varieties of powerful search capabilities, etc
It's got a lot of stuff going on in there.
> ability to use a foreign data wrapper to connect to Redis and build a VIEW out of the result or push data out to Redis/Memcached with a database function,
This sounds amazing. Does anyone have a link to some docs for this?
That "most" is a pretty loaded word! Last I checked, even CREATE INDEX is not part of any ANSI or ISO SQL specification. (That's why every RDBMS has such different features and syntax for this.) Good luck building a system with just "standard SQL".
> even CREATE INDEX is not part of any ANSI or ISO SQL specification
That's correct, and that's why PostgreSQL has ADD CONSTRAINT.
The constraints describe the actual schema of the data; INDEX is an DBMS-specific implementation* for improving performance (including index types, etc.).
* Those since the standard does not yet cover some thing partial unique constraints, these have to be done as INDEX in PostgreSQL.
SQL has nothing to say about the layer of physical implementation, and that's where INDEX belongs (as does STOGROUP [DB2] and what have you).
That's not a bug, that's a feature. Deriving from the deliberate intent to make physical independence (which was simply nonexistent at the time the model was conceived) a reality.
Why should it be? Indexing is an implementation detail for each RDBMS, so it doesn’t make sense to add it to the language spec.
3 replies →
Most of Postgres is standard SQL. It's just that most non-Postgres databases do not implement standard SQL very well.
Sure, but the non-standard enhancements like JSON support are part of what sets Postgres apart from the competition IMO.
SQL:2016 includes support for JSON: https://en.wikipedia.org/wiki/SQL:2016
3 replies →
Wait, what major relational databases don't support JSON? Oracle and Sql Server definitely do. Does MySQL not support JSON?
6 replies →
Doesn't every major RDBMS support JSON out of the box these days?
2 replies →
I feel like Postgres is so powerful I could damn near build an entire web app backend with JUST Postgres (i'm only sorta kidding here). If that's standard SQL, well then I really like standard SQL :-)
I've done pretty much this a few times, and it's amazing if you're working in the context of enterprise data systems that need to provide extensive capabilities to a "small" userbase (i.e. concurrent in the thousands). You just need a small shim layer for security, to transform results sets, and handle browser -> database connectivity.
Postgrest works very well as this shim layer, though I've moved on to writing sql directly in the client and communicating through a web socket shim. In terms of reducing code complexity and improving performance this is absolutely unbeatable, you just need to parse incoming sql to sanitize it and make sure there is no role escalation. Because of postgres's foreign data wrappers this method can provide a consistent surface for basically all your enterprise data. The only gotcha with FDWs is that some of them don't "push down" many query clauses, so you end up doing much slower queries on the remote system and filtering locally, which is terrible for obvious reasons. That being said, the FDWs are pretty much all open source, so you can just implement push down support for those missing clauses yourself.
4 replies →
You think you're kidding.
https://github.com/PostgREST/postgrest
https://wiki.postgresql.org/wiki/Python
https://github.com/plv8/plv8
https://www.graphile.org/
4 replies →
Same. As a data scientist, I have seen multiple 'analytics workbench' solutions come, go and occasionally stay - Clementine (what is now IBM SPSS), Stata, SAS and its many variants, Statistica and in the recent years, tools built off and designed to make working with data using Python/R.
But, the common workhorse tool that has stayed strong through all these has been the common SQL. Elegant, simple, powerful and thoroughly reliable, it is my primary go-to tool. In an otherwise changing ecosystem, its simplicity and reliability is a boon. Yes, it is primarily because of the nostalgic familiarity but I also believe it continues to be extremely powerful, one that will serve you very well.
I came here to write something similar, but you said it better than I could. Sure, SQL has the problems that the article mentions and their solution looks nice on the surface, but SQL has worked much better for me than any alternatives, especially if it’s slightly extended SQL like in Postgres, as you mention, and any attempts to improve it will always be an uphill battle as SQL is pervasive and well supported. That doesn't mean people shouldn't try, but it does mean that their likelihood of success depends on factors other than whether the solution is an improvement over SQL or not.