> The index is 214 MB! That's almost half the size of the entire table. So the analysts are happy, but you? Not so much...
This is part of a broader choice: write amplification. You'd want to, of course, have the most precise index possible - but no matter how you cut it, you are incurring extra I/O for writes - one for the tuple, one per index. How you index things is heavily influenced by the mix of reads and writes, and this is why we have data warehouses/read replicas in the first place: it allows us to avoid write amplification in the write path, while having fast filtered reads (that are slightly delayed).
If you're dealing with <ridiculous number of users>, there is a good chance that you don't want to be putting BI/OLAP indices on your OLTP database. You probably don't have enough users to worry about this - but - if you ever find that your writes are becoming an issue this is something to consider.
Would be nice if PG supported clustered indexes (Index Organized Tables in Oracle speak) as an option if you have a table thats accessed mostly the same way you can get a index without the write amplification because the table is the index.
Clustered indexes only save up to 2x write amplification in the very rare case where you're indexing the entire table (e.g. if it has very few columns).
However, that is usually the least of your concerns with write amplification. If you don't batch your writes, you can easily get 100x write amplification. For any primary key or any other index not strongly correlated with your INSERTs, you can get perhaps another 100x write amplification even if you batch you writes.
>in the very rare case where you're indexing the entire table (e.g. if it has very few columns).
Not sure I follow most tables are accessed primarily in one way (primary key) while maybe sometimes in others for analysis. Having the PK written twice because it's almost always indexed is normally a waste and good candidate for a clustered index. So much so that many DB's like SQLite and MySql always do clustered indexes on primary key because their storage engine is built such that tables are a b-tree anyway vs PG that has separate b-tree indexes and heap tables. MSSQL and Oracle give you a choice whether the table is a index structure or a heap.
If you have very specific use case tables they can typically have a clustered index and no secondary indexes, you can still scan them for ad-hoc analysis but you get better insert performance and space usage because you aren't double writing to the heap and a PK index like you would in PG.
As far as batch writes that is a separate issue and has to due with whether that even makes sense for durability, if you need to commit a single random row due to something occurring you can't batch that up and maintain consistency, if your bulk loading data sure and is common practice to do commit batches there, clustered indexes could still be a 100 vs 200x write amplification if you have to insert both an index row and heap row vs just a single clustered index row.
Another option would be a good way of placing indexes on a different physical disk. You could use fast, ephemeral storage like you can for a WAL without amplifying the writes to the same device that is your expensive bottleneck. You could rebuild on data loss.
But it would add complexity to detect out-of-sync indexes and tables.
If you lose the WAL you lose the data since the last merge but there’s no risk of corruption. The WAL handles missed syncs fine, too, missing losing just that window of data.
I don’t know if or how Postgres records the transaction number in the index to be able to notice if it’s out of date. If it does, I don’t know of any solution to “catch up” the index besides recreating it, which would be ok if that’s the only issue but from my experience with out-of-date indexes (libc or icu updates, where Postgres doesn’t know if anything IS broken and just reports that it could be), there’s no guarantee you’d even notice and your app could be running completely broken until you rebuild.
>After running this command on a data directory with corrupted WAL or a corrupted control file, it should be possible to start the server, but bear in mind that the database might contain inconsistent data due to partially-committed transactions. You should immediately dump your data, run initdb, and restore. After restore, check for inconsistencies and repair as needed.
I guess that depends on the definition of corrupted. It just describes what to do when the WAL is corrupt, but doesn’t say what conditions are considered corrupt. (We use ZFS so I haven’t run into torn writes.)
The most interesting thing for me in this article was the mention of `MERGE` almost in passing at the end.
> I'm not a big fan of using the constraint names in SQL, so to overcome both limitations I'd use MERGE instead:
```
db=# MERGE INTO urls t
USING (VALUES (1000004, 'https://hakibenita.com')) AS s(id, url)
ON t.url = s.url
WHEN MATCHED THEN UPDATE SET id = s.id
WHEN NOT MATCHED THEN INSERT (id, url) VALUES (s.id, s.url);
MERGE 1
```
I use `insert ... on conflict do update ...` all the time to handle upserts, but it seems like merge may be more powerful and able to work in more scenarios. I hadn't heard of it before.
IIRC `MERGE` has been part of SQL for a while, but Postgres opted against adding it for many years because it's syntax is inherently non-atomic within Postgres's MVCC model.
This is somewhat a personal preference, but I would just use `INSERT ... ON CONFLICT` and design my data model around it as much as I can. If I absolutely need the more general features of `MERGE` and _can't_ design an alternative using `INSERT ... ON CONLFICT` then I would take a bit of extra time to ensure I handle `MERGE` edge cases (failures) gracefully.
It's kinda hard to handle MERGE failures gracefully. You generally expect the whole thing to succeed, and the syntax deceptively makes it seem like you can handle all the cases. But because of MVCC, you get these TOCTOU-style spurious constraint violations, yet there's no way to address them on a per-row basis, leading to the entire statement rolling back even for the rows that had no issues. If you are designing for concurrent OLTP workloads against the table, you should probably just avoid MERGE altogether. It's more useful for one-off manual fixups.
That reference - my initial gut feeling was that `MERGE` felt more readable, but then I read this paragraph
> If you want the generality of MERGE, you have to accept the fact that you might get unique constraint violations, when there are concurrent inserts, versus with INSERT ON CONFLICT, the way it's designed with its speculative insertions, guarantees that you either get an INSERT or an UPDATE and that is true even if there are concurrent inserts. You might want to choose INSERT ON CONFLICT if you need the guarantee.
Basically, `MERGE` is susceptible to a concurrent process also writing `INSERT` where that `INSERT` and `MERGE` are unaware of one another, causing a duplicate value to be used.
If you're doing large batch inserts, I've found using the COPY INTO the fastest way, especially if you use the binary data format so there's no overhead on the postgres server side.
Great article, shows a lot of interesting PostgreSQL features. I have used PostgreSQL and MySQL for decades, and this article showed me that I have barely scratched the surface of what is possible.
I've used Postgres for more than a decade and everytime I wade into the docs I feel the same way, I'm barely scratching the surface. It's so immensely powerful.
I love what LLMs are doing for me in PG's SQL. I discovered many features by having LLMs write them for me, often spot-on 100% on first prompt.
Since I know conceptually how RDBMSes work, I can ask vey specifically what I want. Also asking for feedback on schemas/queries really helped me. I use a lot more of PGs features now!
the way PG was originally implemented does have some overlap with operating systems design IMHO.. PG internals define and use PG tables in internal schema to implement core architectural features. The PG code that bootstraps the PG environment is minimal in important ways.
Not discussed in TFA: BRIN indices. For the given date example, they would be perfect - if you can guarantee that whatever you’re indexing is monotonic, they’re a godsend. Tiny and performant.
Minor correction: you don’t even have to guarantee the data is monotonic, it just performs best when it mostly is (good example is when handling incoming timestamps data on the server so messages/rexords are only generally monotonic but may be processed out of order).
How well do they work for UUIDv7? You’d probably have to tune (increase?) pages_per_range, but even though each index entry is 16 bytes you have to consider the btree index on the same is also similarly affected (or worse).
The hash technique for uniqueness isn’t supported for indexes because it doesn’t handle hash collisions. The authors proposed solution suffers the same problem- values which do not already exist in the table will sometimes be rejected because they have the same hash as something that was already saved.
This is completely untrue. While the index only stores the hashes, the table itself stores the full value and postgres requires both the hash and the full value to match before rejecting the new row. Ie. Duplicate hashes are fine.
The no unique constraint using hash index thing is something I always run into. Am I missing something or is it just glue code missing to translate that syntax to a exclusion constraint with a hash index, allowing foreign keys to reference the column in the process, and possibly also fixing the ON CONFLICT approach in one go?
The article explains why they want to avoid this option:
> Starting at version 14, PostgreSQL supports generated columns - these are columns that are automatically populated with an expression when we insert the row. Sounds exactly like what we need but there is a caveat - the result of the expression is materialized - this means additional storage, which is what we were trying to save in the first place!
Thanks, missed that part. I would still be interested in knowing how much additional storage that adds, if the OP is interested in updating the article.
That's the first solution (a function based index), however it has the drawback of fragility: a seemingly innocent change to the query can lead to not matching the index's expression anymore). Which is why the article moves on to generated columns.
It's interesting how both virtual columns and hash indexes work, but feel like they're bolted on, vs being made part of the whole ecosystem so that they work seamlessly.
Virtual columns are basically one or two minor patches from being fully done. Pg18 brought us most of the way there.
Hash indices have long been crippled; they shipped almost unusable but every few years get a good QoL update. I think automatic unique constraints are the big thing left there.
I generally use pgcli to that end. Works well, has a few niceties like clearer transaction state, better reconnect, syntax highlighting, and better autocomplete that works in many more cases than plain psql (it can even autocomplete on clauses when foreign key relations are defined!).
My only gripe with it is its insistence on adding a space after a line break when the query is too long, making copy/paste a pain for long queries.
Sibling comment mentions less space but also your alternative doesn’t (naively) handle supporting values that differ but hash to the same value.
Of course the hash index also outperforms a unique (btree) index on top of separately calculating the hash, in addition to the storage overhead, row bloat, lack of guarantees regarding the hash unless you expose it to Postgres as a user-defined function AND add a check constraint.
>Currently, constraint exclusion is enabled by default only for cases that are often used to implement table partitioning via inheritance trees. Turning it on for all tables imposes extra planning overhead that is quite noticeable on simple queries, and most often will yield no benefit for simple queries.
PG's lack of plan caching strikes again, this sort of thing is not a concern in other DB's that reuse query plans.
PG does reuse plans, but only if you prepare a query and run it more than 5 times on that connection. See plan_cache_mode[0] and the PREPARE docs it links to. This works great on simple queries that run all the time.
It sometimes really stinks on some queries since the generic plan can't "see" the parameter values anymore. E.g. if you have an index on (customer_id, item_id) and run a query where `customer_id = $1 AND item_id = ANY($2)` ($2 is an array parameter), the generic query plan doesn't know how many elements are in the array and can decide to do an elaborate plan like a bitmap index scan instead of a nested loop join. I've seen the generic plan flip-flop in a situation like this and have a >100x load difference.
The plan cache is also per-connection, so you still have to plan a query multiple times. This is another reason why consolidating connections in PG is important.
Yes manual query preparation by client [1] is what you did in MSSQL server up until v7.0 I believe, which was 1998 when it started doing automatic caching based on statement text. I believe it also cached stored procedures before v7.0 which is one reason they were recommended for all application code access to the database back then.
MSSQL server also does parameter sniffing now days and can have multiple plans based on the parameters values it also has a hint to guide or disable sniffing because many times a generic plan is actually better, again something else PG doesn't have, HINTS [2].
PG being process based per connection instead of thread based makes it much more difficult to share plans between connections and it also has no plan serialization ability. Where MSSQL can save plans to xml and they can be loaded on other servers and "frozen" to use that plan if desired, they can also be loaded into plan inspection tools that way as well [3].
This is part of a broader choice: write amplification. You'd want to, of course, have the most precise index possible - but no matter how you cut it, you are incurring extra I/O for writes - one for the tuple, one per index. How you index things is heavily influenced by the mix of reads and writes, and this is why we have data warehouses/read replicas in the first place: it allows us to avoid write amplification in the write path, while having fast filtered reads (that are slightly delayed).
If you're dealing with <ridiculous number of users>, there is a good chance that you don't want to be putting BI/OLAP indices on your OLTP database. You probably don't have enough users to worry about this - but - if you ever find that your writes are becoming an issue this is something to consider.
reply