I think most people think an anti-pattern is an aberration in the "solution" section that creates more problems.
So here, the anti-pattern is that people use a term so casually (e.g., DevOps) that no one knows what it's referring to anymore.
(The problem: need a way to refer to concept(s) in a pithy way. The solution: make up or reuse an existing word/phrase to incorporate the concept(s) by reference so that it can can, unambiguously, be used as a replacement for the longer description. )
Strange choice of example! I'm not sure I agree that your example is a common problem, and I'm even less sure that the proposed solution to it is generally useful.
it isn't, is the thing.
if you read the book design patterns, they spell out what a pattern is.
if you read the book anti-patterns, he spells out what an anti-pattern is.
people have gotten the wrong idea by learning the phrases from casual usage.
but also, the book anti-patterns is pretty clear here
Is this code for 'use a lookup table' or am I falling behind on the terminology? The modern term should be 'sum table' or something similar surely.
'Landed table'? Is that the 'fact table', the one that contains the codes that need to be looked-up?
* in whatever order they're used
if your case statement is just a series of straighahead "WHEN x=this THEN that", you're very lucky.
the nasty case statements are the ones were the when expression sometimes uses different pieces of data and/or the ordering of the statements is important.
Why wouldn’t you store this information in a table and query it when you need it? What if you need to support other languages? With a table you can just add more columns for more languages!
query WHERE name = ‘abc’
create an indexed UPPER(name) column"
Should there be an "or" between these 2 points, or am I missing something? Why create an UPPER index column and not use it?
Unfortunately I learned this the hard way!
Otoh, it seems a fairly stable language (family of dialects?) so finding the pitfalls has long leverage
ALTER TABLE example ADD name_ci AS name COLLATE SQL_Latin1_General_CI_AS;
For example, you define an index on UPPER(name_column), and in your query you can use WHERE UPPER(name_to_search_for) = UPPER(name_column), and it will use the index.
> query WHERE name = ‘ABC’
> create an indexed UPPER(name) column
The point is that the index itself is already on the data with the function applied. So it's not a full scan, the way the original query was.
Of course, in this particular example you just want to use a case-insensitive collation to begin with. But the general concept is valid.
Any time I see DISTINCT in a query I immediately become suspicious that the query author has an incomplete understanding of the data model, a lack of comprehension of set theory, or more likely both.
Though fairly recently I learned that even with all the correct joins in place, sometimes adding a DISTINCT within a CTE can dramatically increase performance. I assume there’s some optimizations the query planner can make when it’s been guaranteed record uniqueness.
Distinct is also easily explained to users, who are probably familiar with Excel’s “remove duplicate rows”.
It can also be great for exploring unfamiliar databases. I ask applicants to find stuff in a database they would never see by scrolling, and you’d be surprised how many don’t find it.
>less verbose
Well…
In any case, it depends. OP nicely guarded himself by writing “overusing”, so at that point his pro-tip is just a tautology and we are in agreement: not every use of DISTINCT is an immediate smell.
SELECT * FROM t1 WHERE EXISTS ( SELECT * FROM t2 WHERE t2.x = t1.x );
SELECT * FROM t1 WHERE x IN ( SELECT x FROM t2 );
SELECT * FROM t1 JOIN ( SELECT DISTINCT x FROM t2 ) s1 USING (x);
You could argue that you could fake a semijoin using
SELECT DISTINCT * FROM t1 JOIN t2 USING (x);
SELECT * FROM t1 JOIN t2 USING (x) GROUP BY t1.*;
SELECT * FROM t1 SEMIJOIN t2 USING (x);
although it creates some extra problems for the join optimizer.
Indeed, along that line, I would say that DISTINCT can be used to convey intent... and doing that in code is important.
- I want to know the zipcodes we have customers in - DISTINCT
- I want to know how many customers we have in each zipcode - aggregates
Can you do the first with the second? Sure.. but the first makes it clear what your goal is.
SOMEWHAT-DISTINCT with a fuzzy threshold would also be useful.
At scale, repeated low-cardinality columns matter a great deal.
SELECT zipcode.rural_urban_code, COUNT(*) AS n_customer FROM customer INNER JOIN zipcode USING(zipcode) GROUP BY 1;
> I immediately become suspicious
All I read from that is, when DISTINCT is used, it's worth taking a look to make sure the person in question understands the data/query; and isn't just "fixing" a broken query with it. That doesn't mean it's wrong, but it's a "smell", a "flag" saying pay attention.
There are self-identifying "senior software engineers" that cannot understand what even an XOR is, even after you draw out the entire truth table, all four rows.
It never used to bug me as a junior dev, but once a peer pointed this out it became impossible for me to ignore.
`if(X&IsFooMask != 0)`
:)
bool x;
...
if (x == true) {
DoThing1();
} else if (x == false) {
DoThing2();
}
When making a code change which touches a lot of places, it's not always obvious to "zoom out" and read the surrounding context to see if the structure of the code can be updated. The developer may be chewing through a grep list of a few dozen locations that need to be changed.
https://hackage.haskell.org/package/base-4.21.0.0/docs/Data-...
There are few other legitimate use cases of the regular `DISTINCT` that I have seen, other than the typical one-off `SELECT DISTINCT(foo) FROM bar`.
I'll test again, really the last time I tested that was two decades ago.
I'm curious, can you demo this?
Do you recall what the database server was?
I also tested this once years later when doing a Python app with sqlite. Similar result, but admittedly that was not a very big table to begin with.
I am meticulous with my database schemas, and periodically review my indexes and covering indexes. I'm no DBA, but I believe that the database is the only real value a codebase has, other than maybe a novel method here and there. So I put care into designing it properly and testing my assumptions.
But if you just have a LIMIT, then no - any RDBMS should stop as soon as it’s reached your requested limit.
"given a BTreeMap<String, Vec<String>>, how do I do .keys() and .len()".
The big justification for its design is to enable compiler optimizations (query planning) but compilers can optimize imperative code very well too, so I wonder if you could get the same benefits with a language that's less declarative.
In fact, IIRC, using DISTINCT (usually bad for performance, btw) is an SQL advice by CJ Date in https://www.oreilly.com/library/view/sql-and-relational/9781...
Oh you looked the schema for t and it said x has a PRIMARY or UNIQUE constraint?
Ah well two minutes after you looked at the schema Tom removed the UNIQUE constraint. Now your scratching your head when you get duplicates.
Sql is a bag language not a set language. The contract with relation t is that if the runtime can find there rel t and attribute x it will return it. You may end up with rows or not, and you may end up with duplicates or not, and the type of x may change between subsequent execution.
So if you want a set you need to say so using DISTINCT. At runtime the query planner will check the schema and if the attribute is UNIQUE or PRIMARY it will not have to do a deduplication.
Certain languages, formats and tools do this correctly by default. For the others you need a source of truth that you generate from.
Though sure, known to negatively affect performance, I think in some database systems more than in others?
> Schema evolution can break your view, which can have downstream effects
Select * is the problem itself in the face of schema evolution and things like name collision.
In sqlite, the view definition will be automatically expanded and one of the columns in the output will automatically be distinguished with an alias. Which column name changes is dependent on the order of tables in the join. This can absolutely break code.
In postgres, the view columns are qualified at definition time so nothing changes immediately. But when the view definition gets updated you will get a failure in the DDL.
In any system, a large column can be added to one of the constituent tables and cause a performance problem. The best advice is to avoid these problems and never use "select *" in production code.
This mirrors how adding additional fields to an object type in a programming language usually isn’t considered a breaking change, but changing the type of an existing field is.
In a better language, this would be a pipeline. Pipelines are conceptually simple but annoying to debug, compared to putting intermediate results in a variable or file. Are there any debuggers that let you look at intermediate results of pipelines without modifying the code?
If you want to build a pipeline and store each intermediate result, most tooling will make that easy for you. E.g. in dbt, just put each subquery in its separate file, and the processing engine will correctly schedule each subresult after the other. Just make sure you have enough storage available, it's not uncommon for intermediate results to be hundreds of times larger than the end result (e.g. when you perform a full table join in the first CTE, and do target filtering in another).
In some languages, a series of assignments and a large expression will often compile to the same thing, but if written as assignments, it will make it easier to set breakpoints.
F# in the visual studio debugger does a pretty good job of this in recent versions.
where a=1
And k=2
And v=3Frankly, that sounds like one of those things that totally makes sense in the author’s head, but inconsiderately creates terrible code ergonomics and needless cognitive load for anyone reading it. You know to just ignore those expressions when you’re reading it because you wrote it and know they have no effect, but to a busy code reviewer, it’s annoying functionless clutter making their job more annoying. “Wait, that should do nothing… but does it actually do something hackish and ‘clever’ that they didn’t comment? Let’s think about this for a minute.” Use an editor with proper formatting capability, and don’t use executable expressions for formatting in code that other people look at.
It means you can copy paste any where condition because they are all of the form
AND <condition>
WHERE <condition 1>
I use this technique for some analytics queries that all rely on the same base table. It’s not uncommon to start with copying an old query and just adding or removing conditions and grouping/aggregating until I get the right data. Using this format also makes commenting out any condition trivial.
WHERE true
-- AND some_column = “some value”
AND event = “SOME_EVENT_TYPE”
AND EXISTS(SELECT * FROM UNNEST(array_column) as v WHERE v = “some value”)
I've seen it used in dozens of places, in particular places that programmatically generate the AND parts of queries. I wasn't really that confused the first time I saw it and I was never confused any time after that.
No, you ask the DB to EXPLAIN itself to you.
Translating status codes into English or some other natural language? That's better done in the application, not the database. Maybe even leave it to the frontend if you have one. As a rule of thumb, any transformation that does not affect which rows are returned can be applied in another layer after those rows have been returned. Just because you know SQL doesn't mean you have to do everything in SQL.
Deeply nested subqueries? You might want to split that up into simpler queries. There's nothing shameful about throwing three stones to kill three birds, as long as you don't fall into the 1+N pattern. Whoever has to maintain your code will thank you for not trying to be too clever.
Also, a series of simple queries often run faster than a single large query, because there's a limit to how well the query planner can optimize an excessively complicated statement. With proper use of transactions, you shouldn't have to worry about the data changing under your feet as you make these queries.
I wrote a small tutorial (~9000 words in two parts) on how to design complicated queries so that they don't need DISTINCT and are basically correct by construction.
https://kb.databasedesignbook.com/posts/systematic-design-of...
Edit: it’s also actually a book!
Using != or NOT IN (...) is almost always going to be inefficient (but can be OK if other predicates have narrowed down the result set already).
Also, understand how your DB handles nulls. Are nulls and empty strings the same? Does null == null? Not all databases do this the same way.
Also in regards to indexing. The DBs I've used have not indexed nulls, so a "WHERE col IS NULL" is inefficient even though "col" is indexed.
If that is the case and you really need it, have a computed column with a char(1) or bit indicating if "col" is NULL or not, and index that.
If your business rules say that "not applicable" or "no entry" is a value, store a value that indicates that, don't use NULL.
I guess you would handle it in the application and not in the query, right?
If you have a table of customers and someone of them don't have addresses, it's standard to leave the address fields NULL. If some of them don't belong to a company, it's standard to leave the company_id field NULL.
This is literally what NULL is for. It's a special value precisely because missing data or a N/A field is so common.
If you're suggesting mandatory additional has_address and has_customer_id fields, I would disagree. You'd be reinventing a database tool that already exists precisely for that purpose.
Kinda. You need null for outer joins, but you could have a relational DBMS that prohibits nullable columns in tables. Christopher Date thought that in properly normalised designs, tables should never use nullable columns. Codd disagreed. [0]
> If you're suggesting mandatory additional has_address and has_customer_id fields, I would disagree. You'd be reinventing a database tool that already exists precisely for that purpose.
The way to do it without using a nullable column is to introduce another table for the 'optional' data, and use a left outer join.
[0] https://en.wikipedia.org/wiki/First_normal_form#Christopher_...
I mean, you could, but having separate tables for every optional field would be an organizational and usability nightmare. Queries would be longer and slower for no good reason. Not to mention a gigantic waste of space with all those repeated primary keys and their indexes.
And you could have databases that prohibited NULL values, but we mostly don't, because they're so useful.
I think this indicates that declaring and managing state is too onerous in SQL.
The idea that having a separate table for every optional field is too unworkable isn't an issue with SQL. It's a fundamentally overcomplicated idea. It's like a programming language where every variable that could be null had to be put in its own file. It would be terrible design.
I remember working on ERP systems with 40+ column tables, most of which were null. With no clear constraints on which options should or shouldn’t enable or make mandatory other options. This becomes incredibly obvious and natural when you group in additional tables.
My tables are incredibly concise and the cache loves this.
Having a bunch of grouped optional values in other tables adds a ton of maintenance and query complexity.
The "clear constraints" belong in business logic, whether in triggers at the database level or before queries get executed at the application level.
Now, splitting up tables can produce performance optimizations. I'm not saying to never do it. But it's a tradeoff that increases complexity.
Those are rookie numbers. Add a zero to that number and we're talking.
And for us, a good portion of the data, a considerable fraction of those fields will have data, and which fields will vary between customers.
All except some key fields are NULL-able since the user can save and resume their work.
Just to display our main screen would require 100-150 joins using a separate table per optional.
I'm pretty sure the database would not love that.
table Entity_Fields (ParentEntityId : int not null, FieldId : int not null, IntValue : int, TextValue : varchar(MAX), DateValue : datetime, ...)
table Entity_IntFields (ParentEntityId : int not null, FieldId : int not null, Value : int not null)
table Entity_TextFields (ParentEntityId : int not null, FieldId : int not null, Value : varchar(MAX) not null)
table Entity_DateFields (ParentEntityId : int not null, FieldId : int not null, Value : datetime not null)Could have been, but no. And doing it like you suggest would mean overview grids would have to do 50+ subqueries for each row, and loading a record would mean hundreds of queries.
And insertion performance would crater I assume, since the DB now needs to do hundreds of inserts per record rather than a single row. We do have some cases where we get 100k inserts per hour once a day or so. And this is just the main table, we have many child tables already, though they're not nearly as wide.
I think a more realistic split would have been to split the main table into maybe 10 tables or so at most. Still would result in a fair bit of subqueries for grids and such, but not that bad.
Only if you insist on loading the data into a flat record type with all of those fields. You have a dynamically evolving datatype, so your class should also be dynamic:
public enum FieldId
{
Field1,
Field2,
//...
}
public class YourRecord
{
public Dictionary<FieldId, string> TextValues { get; set; }
public Dictionary<FieldId, int> IntValues { get; set; }
public Dictionary<FieldId, DateTime> DateValues { get; set; }
// ...
}
> And insertion performance would crater I assume, since the DB now needs to do hundreds of inserts per record rather than a single row. We do have some cases where we get 100k inserts per hour once a day or so.
That's a fairly small amount of data, I don't really see an issue here.
Also, you're neglecting the fact that since the fields are independent, adding a field value doesn't involve locking and rewriting a large 400 column row to disk, but only appending a tiny 4 column row to a separate distinct table.
Fair enough. Application logic will want to access the data as a flat record, but that could be handled through getters. We have views which are also used by customers and our reporting tool, but the customers we're moving to API access and reporting could probably be done with something better if starting from scratch.
> That's a fairly small amount of data, I don't really see an issue here.
Well it turns 100k inserts into tens of millions. On tables where users need to work without slowdown while this is going on.
It probably works if you throw enough hardware at it, but currently we get by with quite modest DB hardware.
That said, how do you create compound indexes over these fields? Say you need to index one date and one varchar column? Such demands can arise down the line, often hard to predict up front.
I don't think "number of inserts" is the right metric because the total amount of bytes being written is almost the same, it's just written in different areas and still mostly contiguously. I think "number of distinct tables being written" is a better metric. Assuming all 400 columns become 400 records in the data type tables, say evenly divided among the 4 most common data types (int, decimal, text, date), that would be more like (4 or 5) x 100k = 400k-500k. I would still hesitate to naively compare it this way without a benchmark though, because with 4 or 5 tables being written there's also less contention than there is when everyone is writing to 1 table.
Regarding indexes, you can index the field tables but obviously this applies to the whole table. Without more understanding of your domain I can't really say if this breaks down. There's also the possibility that this one table is serving two competing goals, eg. perhaps the user should be able to add data incrementally (so this pattern applies), but then at some point when all of the data is filled it should be migrated to an actual table with all non-null columns that can be indexed and queried/processed as you usually do.
In any case, what I've sketched out isn't really new, it's been used at least since the 90s under "Dynamic Object Model" and "Entity-Attribute-Value" [1,2], so if they were using on hardware in the 90s I can't imagine the pattern would be unusable on modern hardware.
[1] https://www.cs.sjsu.edu/~pearce/oom/patterns/new/Riehle.pdf
The answer, obviously, is because traditional tables make lots of things really easy. SQL is designed for that use case, and it's performant.
You're not even talking about traditional relational databases anymore. You're trying to construct tables inside of tables, which means abandoning performance, indexes, etc.
Because there is a schema, it's just a dynamically evolving one, and also, presumably the rest of your system depends on the relational DB, so why multiply your dependencies unnecessarily?
> You're not even talking about traditional relational databases anymore.
There's nothing non-relational about the schema I've outlined, it matches what the domain requires with less noise than 400 nullable columns.
It sure is. Consider a database language that innately supported algebraic data types. Instead of:
table Contact { int Id; Nullable<string> email; Nullable<string> phoneNo; }
type PhoneOrEmail = Phone<string> | Email<string>
table Contact { int Id; PhoneOrEmail info; }
No, these aren't SQL limitations, it's design that is super complex. Figuring out how to index over these multi-type fields isn't a SQL limitation, it's a hard engineering problem.
Actually that's exactly what you'd want, because it saves you from running two different queries in those cases with properly normalized disjoint data sets, and moves more of the domain's constraints into the database schema where it belongs.
> and moves more of the domain's constraints into the database schema where it belongs.
No, I prefer to keep my domain constraints at the application level where they're far more flexible.
The database is for storing information, not for validating my business logic.
I mean, I realize some people want to build some of that logic into the database, especially when many applications interact with it. But it's not a superior design pattern. If you have a single application, it's perfectly valid and desirable to put all business logic in the application, not the database.
https://www.postgresql.org/docs/current/tutorial-inheritance...
Agreed that queries would tend be longer as you'd need joins, although views could help, especially for read operations.
Regarding storage-efficiency and query speed, agreed that it could well hurt both, but it's going to depend. If a column holds null in almost all rows, we would expect it to be more space-efficient to use a separate table and a left outer join. Query speed could also improve for queries that don't reference the nullable column, as the 'main' table would be smaller in storage. (I'm assuming a rowstore here.)
If I have a column for the ID of the customer's current active subscription, and that column is NULL, it seems perfectly fine to interpret that the customer has no active subscription.
That is a valid inference. You don't need a separate has_active_subscription field.
On the other hand, your phone number example is just common sense. The database doesn't represent the external world. The database just knows the customer didn't provide a phone number.
Why do you say that?
My understanding is that as long as the RHS of NOT IN is constant (in the sense that it doesn't depend on the row) the condition is basically a hash table lookup, which is typically efficient if the lookup table is not massive.
What's the more efficient alternative?
If I have a table of several million rows and I want to find rows "WHERE foo NOT IN ('A', 'B', 'C')" that's a full table scan, or possibly an index scan if foo is indexed, unless there are other conditions that narrow it down.
The biggest problem with NOT IN is that it has very surprising NULL behavior: Due to the way it's defined, if there is any NULL in the joined-on columns, then _all_ rows must pass. If the column is non-nullable, then sure, you can convert it into an antijoin and optimize it together with the rest of the join tree. If not, it usually ends up being something more complicated.
For this reason, NOT EXISTS should usually be preferred. The syntax sucks, but it's much easier to rewrite to antijoin.
Note that these caveats do _not_ apply to IN, only NOT IN.
Therefore you should create a consistent indentation style for SQL. See https://bentilly.blogspot.com/2011/02/sql-formatting-style.h... for mine. Second, you should try to group logical things together. This is why people should move subqueries into common table expressions. And finally, don't be afraid of commenting wisely.
How that auto-formatter indents is borderly almost a hate crime. A thousand times better to indent manually.
If anyone wants to check out a half-done lang with lacking documentation, I'd be happy to read your feedback: https://lutra-lang.org
Experts including Codd recognized the problems with SQL since that language got traction. Some alternatives got proposed, perhaps most notably Tutorial D by Chris Date and Hugh Darwen. No SQL replacement goes anywhere because of the vast quantity of SQL code and supporting tools dating back decades. Chris Date wrote the textbook on databases, and at least one book going through the problems with SQL and various implementations of the relational model.
SQL perfectly illustrates what Strostrup meant by "There are only two kinds of languages: the ones people complain about and the ones nobody uses." In some sense I would welcome a better query language. On the other hand I attribute decades of job security and steady income to knowing SQL and dealing with its problems.
* Don't store UUIDs as strings.
* Don't use random UUID variants for your primary key (or don't use UUIDs for your primary key).
* Don't use a random column in your clustered index.
This is still 2x the space of an auto increment number.
This is overhead for every table, every index, and every relationship.
That might be acceptable in your case though, the case where it became unacceptable in my experience was in a MSSQL Express context. But it was an idiotic decision to use MSSQL to begin with in that scenario.
Regarding random clustered indexes. Broadly speaking you want your clustered index to be made up of some incremental unique set of fields.
I mean, technically there is not a massive issue, but the largest tables in your database will be the non-indexes (indexes are just tables) and you want your big, mainly append only, tables to be nicely compact so a bunch of space isn't taken up by half full pages.
But again, I should honestly have clarified that the problem was mainly an MSSQL Express problem where databases are limited to 10GiB.
You might honestly be fine, but do look for documentation on your specific database.
User Defined Functions (UDFs) are another option to consolidate the logic in one place.
> Using Functions on Indexed Columns
In other words, the query is not sargable [0]
> Overusing DISTINCT to “Fix” Duplicates
Orthogonal to author's point about dealing with fanout from joins, I'm a fan of using something like this for 'de-duping' records that aren't exact matches in order to conform the output to the table grain:
ROW_NUMBER() OVER (PARTITION BY <grain> ORDER BY <deterministic sort>) = 1
[0] https://en.wikipedia.org/wiki/Sargable
[1] https://docs.aws.amazon.com/redshift/latest/dg/r_QUALIFY_cla...
The funny thing is it's actually several of those languages. :-)
was surprised to not see anything about dates/time.
Very often I have seen this problem buried in code design and it always sucks. Sometimes an orm obscures this but the basic antipattern looks like
Select some stuff
For each row in stuff:
… do some important things …
Select a thing to do with this row
… maybe do some other things …
The non-sucky version of the code above is
Select some stuff, joining on all the things you need for the rows because databases are great
For each row in stuff:
… do some important things …
… maybe do some other things …Many materialized views that rely on materialized views. When one at the bottom, or a table, needs a changed all views need to be dropped and recreated.
Using a warm standby for production. I love having a read only production database, but since it's not the primary, it always feels like it's on the losing end of the system. Recently upgraded to Postgres 18 and forgot that means I need to rm rf the standby and pg_basebackup to rebuild... That wasn't fun.
Your code should handle the data model and never allow bad states to enter the database.
There's too much performance loss and too many footguns from these "features".
Our staging environment has its own instance that is rebuilt from prod, with pii removed, every day outside working hours (this normally takes about 15 minutes). It’s fantastic for testing migrations, and is easy to support compared with a warm standby.
I switched to warm standby to reduce stress on the production db which was in the cloud. There is just a single production server and having it constantly run the heavy data processing MVs + handle queries was CPU intensive and slowed everything down. The CPU was costly.
To fix those issues, especially the CPU, I run the primary on a home server where it can crank the CPU as much as it wants running the data processing MVs and then sends the processed WALs to the warm standby that just handles the queries.
This has fixed those CPU and slow queries (when an MV is updating a table that is being constantly read). But introduced headaches anytime I update postgres.
My understanding is the 'fix' is to move data processing to another postgresql DB or flow? My biggest reason for not using another DB is I didn't like the idea of losing direct relations for keys.
Anyways, I appreciate the input, it's been a thorny issue I hit once or twice a year and am always unsure if what I'm doing is 'normal' or what I should do to fix it.
https://en.wikipedia.org/wiki/Sargable
https://www.brentozar.com/blitzcache/non-sargable-predicates...
And Google explains "The term 'sargable' is a portmanteau of "Search ARGument ABLE," formed by combining the words from a SQL database context."
I've been working with SQL for 20+ years, and have literally never come across that word a single time in any documentation, tutorial, Stack Overflow answer, or here on HN. Working in Postgres, MySQL and SQLite.
Is it used at some particular company, or open source community, or with a particular database, or something?
Consider the Simple example it presents. The article is in effect implying that no query optimiser would be able to figure out the equivalence of the two predicates.
(Let's ignore that the two predicates aren't actually equivalent; the first version may raise an exception if myIntColumn is negative, depending on the DBMS.)
Ozar's article is much better. It doesn't make sweeping assumptions about the limitations of all query optimisers, or basic oversights in contrasting supposedly equivalent predicates.
Guilty as charged. I love to do this. Materialized views aren't really possible on sqlite, and so I find stacking views on top of one another very readable and manageable. But it's true other people find it a little obscure and weird.
Some of these have nothing to do with SQL the language itself, and more to do with database schema design. If you have to do a DISTINCT, it means your primary key design is likely not right. If you are layering too many views, something is broken in the base table design, requiring the creation of all these views.
A good database model goes a long way to avoiding all this.
1. No easy way to limit child records count in joins - find all orders with orderproduct.amount is greater than X. Obviously this will genereate duplicates for orders that have more than one such orderproduct. So you slap a distinct on it… but what if you need an aggregation?
The possible fixes are highly non-trivial: subqueries, window functions, or vendor specific: outer apply.
2. Or queries, that is when you group where conditions with OR are very hard (impossible) to optimize.
Apart from the trivial case where the conditions are all on the same column, you are better of leaving the declarative world and imperatively tell sql to do a union.
I wrote a bit about it here: https://www.inuko.net/blog/platform_sql_or_conditions_on_joi...
in 2, looking at your article, from your first query it looks like person_relationship contains both (A,B) and (B,A) for all related people A and B; otherwise the left join won't work. If you also make people related to themselves and store (A,A) and (B,B) there your query becomes much simpler:
SELECT other.id, other.name
FROM person p
JOIN person_relationship r ON r.from_person_id = p.id
JOIN person other ON r.to_person_id = other.id
WHERE p.family_id = @familyId;If I have to work with one more "custom field" or "classifier" implementation, I am going to cry. Your business domain isn't too hard to model, if you need a 100 different "entities" as a part of it, then you should have at least 100 different tables, instead of putting everything into an ill fitting grab bag. Otherwise you can't figure out what is connected to what by just looking at the foreign keys pointing to and from a table, because those simply don't exist. Developers inevitably end up creating shitty polymorphic links with similarly inevitable data integrity issues and also end up coupling the schema to the back end, so you don't get like "table" and "table_id" but rather "section" and "entity_id" so you can't read the schema without reading the back end code either. Before you know it, you're not working with the business domain directly, but it's all "custom fields this" and "custom fields that" and people end up tacking on additional logic, like custom_field_uses, custom_field_use_ids, custom_field_periods, custom_field_sources and god knows what else. If I wanted to work with fields that much, I'd go and work on a farm. Oh, you're afraid of creating 100 tables? Use codegen, even your LLM of choice has no issues with that. Oh, you're too afraid that you're gonna need to do blanket changes across them and will forget something? Surely you're not above a basic ADR, literally putting a Markdown file in a folder in the repo. Oh, you're afraid that something will go wrong in those 100 migrations? How is that any different than you building literally most of your app around a small collection of tables and having fewer migrations that will affect pretty much everything? Don't even get me started on what it's like when the data integrity issues and refactoring gone bad starts. Worst of all, people love taking that pattern and putting it literally everywhere, feels like I'm taking crazy pills and nobody seems to have an issue what it's like when most of the logic in your app has something to do with CustomFieldService.
Fuck EAV/OTLT, thanks for coming to my rant. When it comes to bad patterns, it's very much up there, alongside using JSON in a relational database for the data that you can model and predict and put into regular columns, instead of just using JSON for highly dynamic data.
> Excessive View Layer Stacking
> In larger data environments, it’s easy to fall into the trap of layering views on top of views. At first, this seems modular and organized. But over time, as more teams build their own transformations on top of existing views, the dependency chain becomes unmanageable. Performance slows down because the database has to expand multiple layers of logic each time, and debugging turns into an archaeological dig through nested queries. The fix is to flatten transformations periodically and materialize heavy logic into clean, well-defined base views or tables.
I will say that this is nice to strive for, but at the same time, I much prefer having at least a lot of views instead of dynamically generated SQL by the application (see: myBatis XML mappers), because otherwise with complex logic it's impossible to predict exactly how your application will query the DB and you'll need to run the app locally with logging debug levels on so you see the actual SQL, but god forbid you have noisy DB querying or an N+1 problem somewhere, log spam for days, so unpleasant to work with. It's even more fun when people start nesting mappers and fucking around with aliases, just give me MongoDB at this point, it's web scale.
jwsteigerwalt•3mo ago