▲ 523 points
back
240 comments
Some comments and corrections:
* Use uuidv7 not uuid in general (typically v4)
* in addition to minimizing locked records, make sure your locks are ordered deterministically across all queries (eg by id asc, always) or you’ll deadlock (but postgres has a really good deadlock detector so you’ll more likely just error out if you’re lucky)
* always use explain (generic_plan) to be able to a) copy-and-paste your queries with placeholders for parameters as-is, b) see how your query will actually be optimized when Postgres doesn’t have visibility into the specific parameter values
* use set seqscan = off when testing your query plans esp when tables are empty or nearly so so you can see if indexes will be used when seq scans become less cheap
* everyone defaults to btree indexes which are heavy and increase index bloat. Consider using a hash index instead if you just need to look up by column/id but not sort or get values greater/lesser than a param. You can’t create unique hash indexes but you can create exclude using hash constraints for the same effect (except no multicolumn unique index support)
* learn about GIN (and GIST) indexes. They can speed up common queries without needing new syntax, something people coming from MySQL might not expect to be possible; i.e. you can use them to speed up Plain Jane like ‘%foo%’ queries without switching to FTS.
OP here, I appreciate this.
> in addition to minimizing locked records, make sure your locks are ordered deterministically across all queries (eg by id asc, always) or you’ll deadlock (but postgres has a really good deadlock detector so you’ll more likely just error out if you’re lucky)
This is really good advice, I should put this somewhere in the guide. To add to this, not only can you deadlock by not having a consistent `ORDER BY` when you're locking sets of rows, but you should also be careful of locking rows on tables in different orders. For example, even if you lock each row in a table with an ORDER BY and FOR UPDATE, if one tx locks `table_a` and then `table_b`, and the other locks `table_b` and then `table_a`, you'll deadlock. This is obvious in theory but exponentially harder to debug in practice, because you need to be globally aware of every table that a write touches - something that's bitten us in particular with certain extensions.
> learn about GIN (and GIST) indexes
We're just testing GIN for fast key-value lookups for JSONB columns, and the performance improvements have been really massive. Interestingly there was a large performance skew between AND vs OR on these key-value queries.
Any kind of uuid PK is quite expensive and usually not worth, because you're so frequently joining on PKs. A safe default is to use serial PKs, then have a secondary-indexed uuid4 if you wish to publicly-expose anything. Why uuid7, is the btree performance better with it than with uuid4?
> Use uuidv7 not uuid in general (typically v4)
7/4 'converters' have been featured on HN a few times:
> use set seqscan = off when testing your query plans esp when tables are empty or nearly so so you can see if indexes will be used when seq scans become less cheap
How well does this work for you? I thought if you have _any_ index, Postgres will use it if you disabled sequential scans. Diabling sequential scans won't tell if you if you have the right index
good advice!
> learn about GIN (and GIST) indexes
yes, but also learn about the trade-offs and in particular the "pending list". The flush of the pending list can be slow, causing timeout, causing the list to not be flushed, causing the next write to trigger flush again and failing in the same way, which means you're having downtime. The default pending list size is weirdly high IMO
Should one of the first things you do with a database not be to have a backup strategy? I understand that HA would be a "nice to have" when first starting out, but surly if you have a production db, a backup and restore plan should be on a survival guide? Neither appear to be mentioned here.
What do you all use for your pg backups? Is Barman ( https://pgbarman.org/) still the way many do it? (I haven't deployed a new pg instance for a while, but thinking about it for a new project).
I might get flak for saying this but if you aren't a postgres expert already: just use RDS or a similar cloud DB. The amount of money you're saving by hosting and managing your own postgres instance is absolute peanuts compared to having battle-tested infrastructure for HA, backup and restores, point-in-time recovery, read replicas, etc.
FWIW, we use: https://pgbackrest.org/
Offers point-in-time recovery which is an improvement over a custom solution we used to have which gave us nightly backups.
We have it backing up to Backblaze B2 (S3 like). Was relatively easy to setup and no problems really.
There’s no need to get all complicated and fancy or introduce more dependencies. For most people, a cron job calling pg_dump_all piped to zstd and copying the output to s3/ftp/whatever is plenty good enough.
Obviously past a certain point carting around full backups becomes time/dollar prohibitive, but this can take you very far.
An atomic volume snapshot should work for any database that's durable on power failure. Ideally preceded by a checkpoint, to minimize recovery time. Atomicity of the snapshot mechanism is essential to prevent data corruption using this approach.
We used EBS snapshots on AWS for multi-TB MongoDB to get incremental backups that are fast to create and fast to restore (with some performance degradation after restore).
It doesn't support point-in-time recovery, but since it's fast you can create frequent snapshots (e.g. hourly). I'd consider adding this as a secondary backup strategy, even if you use a higher-level postgres-specific backup tool.
If you already run k8s, why not just use cnpg?
This advice is good, but every startup I've worked with has run into lower hanging fruit than this even. Less scaling problems and more just organizational. Usually what fixes that is:
1. Don't use an ORM.
2. Use serial PKs, not meaningful fields (article mentions this).
3. Use jsonb if needed, but sparingly.
4. Make your source of truth append-only, meaning you only insert, never update or delete. You can have secondary denormalized tables that are mutated, but that's only for performance/convenience and shouldn't be your sot.
5. Use connection pools, but be mindful of how many connections you're using. You probably don't need PgBouncer unless you've messed something up.
6. In code, avoid explicit transactions unless there's a clear reason you need them. Usually only need those for denormalized parts. Just take a conn from the pool, do something, commit, return conn to pool. If you're going to keep an xact open, never do long-running stuff in the middle like RPCs. Too often I see people leave xacts open without much thought. Edit: Also don't use SERIALIZABLE xacts almost ever.
7. Something is probably wrong if you're using explicit locking like SELECT FOR UPDATE.
8. Don't reinvent a type system by having a single table where each row can mean many different things depending on a "type int" enum col. Seems oddly specific, but for some reason someone always tries this.
9. Related to above, don't reinvent a graph DB, typically with "node"/"edge" tables that FK into themselves or in a cycle. 99% of the time what you're trying to do is easily solvable with regular normalized tables.
Having been early at a startup that relied on Postgres I think this post doesn't put enough focus on monitoring and alerting. Postgres has a few key failure modes that you want to avoid ever happening, and you can use alerting to get early warning that you're danger of it happening.
For example, AWS will send you an email if you're approaching XID wraparound. In a startup that email is very likely to be missed, especially if it's sent on Boxing day. You want whatever AWS is watching to send you that email to be something connected to a pager.
Good article overall, some comments:
> Use foreign keys with cascading deletes for low-volume tables, particularly where database consistency and correctness are important. Careful at higher volume.
This might be just me, but I hate cascades, for a very simple reason: at most places, the majority of developers "live" in the Python/Node/Go/whatever application that talks to the database, not the database itself. Cascading deletes (or updates) is basically magic and it can be very hard to understand "why did deleting a row from table A delete something from table B automatically". Especially if someone sets up the cascading wrong! IMO it's better for long-term maintainability to emit explicit delete clauses. Correct use of foreign keys will prevent any issues with database consistency.
> Tricks for large table migrations
The pitfalls and workarounds are all correct, but worth pointing out tooling already exists[0] for managing this for you. Making changes to large tables should be as simple as running a command (and then nervously monitoring for the next 24 hours as the data copies).
Other things to consider,
1. Get used to separating application and database deployments early. It is impossible to transactionally deploy both a schema change and an application change simultaneously, there will always be some delay where the versions of database and application are out of sync, and you will eventually run into a situation where the database change deploys fine but your application change does not. Once your app is in production, get in the habit of only doing backwards compatible schema changes: all new columns are nullable or have a default, no renaming of tables/columns, etc.
2. In the same vein, figure out a schema management strategy early. You really don't want your database deployment process to be "senior dev runs some DDL manually on production from his machine". I'm still partial to liquibase because it's the devil I know, but there's other tooling like Flyway which exists.
(Matt from Hatchet)
One small addendum here is we've had a lot of success performing joins in memory in a few very specific situations where the alternative is a single, often overcomplicated query. I've heard / seen advice many times in the past about performing fewer round trips to the database being something to optimize for (often good advice!). Sometimes this is taken too far, resulting in overly-complex queries requiring complicated JOIN or UNION logic, CASE logic, and so on.
We have a couple of places in our codebase where we perform two or more simpler queries independently instead, and then loop through their results and use maps to match the relevant rows. Conventional wisdom often suggests this path will hurt performance because of the extra database round trip in addition to the loops needed to perform the join, but it is actually beneficial in these cases because of more predictable query planning behavior. We use this trick sparingly, but it can be helpful in a pinch.
Note that some ORMs will also do this for you in the background, which we don't necessarily endorse, and we try to use this sparingly when writing a single query on its own is not realistic.
> Because of all these connection footguns, external connection poolers like pgbouncer are great! If you can’t add this for whatever reason, in-memory connection poolers are a great second option. For example, because Hatchet is open-source, we don’t assume that all user databases use connection poolers, so we use pgxpool (an in-memory connection pool for Go) for this purpose.
Few people know that there is a major bifurcation when it comes to connection pooling implementation.
1. Most application connection poolers follow a first-in-first-out (FIFO) algorithm, which is simple enough to implement and is enough to make sure the application always has a connection available to connect to the database. It optimizes low latency, and works great from the point of view from the application. The problem is that it has few mechanisms to remove redundant connections, since the application is constantly keeping them all "warm".
2. PgBouncer and very few external poolers follow the inverse idea – last-in-first-out (LIFO), and they optimize for reducing the number of connections that reach Postgres, thus improving its throughput. The idea might seem crazy at first – the last connection used is the first one to be picked up again – but this algorithm automatically removes excess connections, which will get cold and get closed.
When starting a new application, option (1) is enough, but as it scales up enough, at some time it is recommended to use (2), since having hundreds of open connections to Postgres is bad for performance if you can use PgBouncer or similar to cut it by 90%. Postgres' process-per-connection design works much better when there are fewer connections reaching it.
To this articles credit, it does start out with normalization and design!
There needs to be more emphasis how important this is! I cant tell you how often I see it done "badly" (we let our ORM build the db for us). The best text I have ever found on this is "Database Design for Mere Mortals", over the years I have bought more that a few copies and I always end up giving them away to those in need (and there are always people around in pretty dire need).
The one thing I would say is missing from this article is to not be afraid of using postgres for "stupid" things. Cache, queue's, and so on, especially on the road to launch.
One should also not be afraid of having more than one Postgres instance, especially if you're using it as a work queue.
Lastly there is a stupid amount of power in Postgres roles (its "user" system). The manual here is somewhat OK, but really undersells richness that it makes available to you.
Postgres is my favourite thing, but I find it's prohibitively costly when bootstrapping something that is lean and frugal.
I end up with a mixture of serverless storage like DynamoDB, S3, DuckDB on S3, and SQLite.
Am I crazy? How can one have a decent Postgres and not pay at least $100/mo (yes, when I say frugal I mean really frugal ... think solo founder that likes to stay on free tiers haha) -- I am aware of Neon/Supabase, but last time I tried them they ended up becoming a tightly coupled annoying dependency after scale that defeated the cost savings as they grew in costs and we ended up migrating to Aurora / RDS lol
EDIT: I'm aware of the self-hosted path but I find configuring the above things faster/cheaper in terms of my admin hours than the self hosted postgres db. Maybe I just suck at being a DBA or need better education on it, that said, I have AI now so I should give it a chance again as it's been a minute since I created a fresh thing
I disagree with the timestamptz advice. I tend to use timestamp (without the timezone), this forces me to use UTC everywhere, so I'm not even tempted to use anything else. I work in fintech, and so far whenever i saw someone storing datetimes with an associated time zone, it always ended with a disaster.
> FOR UPDATE SKIP LOCKED
> The best way to think about this Postgres feature is that it reserves the rows that you’re selecting for use in your transaction without interfering with other queries. We use it primarily for implementing our job queue;
SKIP LOCKED is useful for implementing job queues with interactive transactions – you lock the row while working on it in the application and keeping the transaction open. For high-performance applications it is best to avoid interactive transactions at all, and just update the rows to "pending" immediately. There is no need for SKIP LOCKED in this case.
As a rule of thumb, as you scale up the application, you want to have less state in the database memory, and interactive transactions are just that. Idempotence beats atomicity at scale.
Do folks have any thoughts on ways of avoiding deadlocking access patterns? In a codebase where folks are sort of adding ad-hoc endpoints left and right, it's hard to avoid the case of two endpoints that more or less want to do:
tx1: update a
tx2: update b
tx1: update b
tx2: update a
This guide is nicely formatted and very helpful but I think the comments prove that all it has done is taken a subset of the information from the manual and decided that "these parts are important" whereas the comments then unhelpfully point out, "yes but also this", "and this".
There is no line that says X is important for a startup and Y isn't, it is mostly a matter of degree.
Instead, what the guide has done although could be a little clearer is e.g. explain that there are types of indexes that are a better trade-off of space and performance for particular scenarios, here is one example and click here to learn about the other indexes. That is enough for a startup to understand.
A second example might be, "Optimizing your postgres resource limits is important for x, y, and z reasons. You should try and balance giving your server as much RAM as it can but without using up RAM that is needed by other things. A starting point might be X percent of total RAM for shared_buffers but for more details see the main docs here".
I love postgres but it is not a toy and it takes investment of time. I think what most people want is a map with a few examples so they can pick out what concerns them the most.
> The reason I think it’s useful to view queries as binary—they either seq scan or they don’t seq scan—is: the more you micro-optimize a query, the more of a risk you take that the query planner goes rogue. If you stick to querying by primary keys and indexes, the query planner will have a much easier time.
It's also important to notice the query planner optimizes for the average case, but often it would be better for the app developer if it was optimized for the worst case. But optimizing for the former is a much more tractable problem, so no wonder that's what is implemented.
I had to fight against the query planner when it would optimize a query for the average user, with few rows in a given table, and it would pick one index that made sense for that situation and return a result in less than 10ms. However, when a heavy user issued the same query, depending on the exact parameters the worst case could take over 1 second. So I had to write a much more complex query to force it to take another path with a different index, which would be slower in the average case, but in the worst case would take still less than 100ms. Avoiding timeouts was much more important for my company than taking 10ms more in the average case.
* BRIN indexes are great for append only with an incremental value (a 50MB instead of 100GB index in timeseries data in my case)
* If your disks are ssd and scsi in different volumes, adapt the random cost of io in the config to let the planner now
* if you have RAID controller and a Battery-Backed Write Cache (BBWC), you can disable Linux filesystem write barriers. removing excessive fsync from the mouth of the psql demigod in SoCaL ~linux 2015, Bruce Momjian IIRC
* monitor disk usage, in backups pipe to gz, never to disk
* counter-intuitevely modern hardware may have an io bottleneck and plenty of cpu, so try Filesystems like ZFS using Zstandard (zstd), for boost in your Transactions per second
* if possible, schema multitenancy instead of database multitenancy, instagram talked about this decades ago
* indexes index functions results too, precompute those fields and partial indexes help a lot
* BM25 and FTS in pg are so good you probably don't need Elasticsearch and you will save a lot in de-sync between both of them
* you may be hacked in this brave new world post Mythos, thus, learn PITR to an external only write, no override medium like S3
My advice:
- Don't use long-running transactions. They are a risk for db health. Only use transaction when you have a strong justification
- Set idle_in_transaction_session_timeout to prevent a long-running transaction from holding on to locks or tuples
- Set lock_timeout for migrations to prevent a single DDL statement to bringing down your system
- Set statement_timeout to prevent an expensive query from bringing down your system
On migrations, there's a .Net tool called Grate that I tend to use for schema migrations... I don't use all the features, but it works well... using a migration stack in a repository for deployments and a similar tool is IMO more reliable than magic comparison tools or hand migrations in practice. You should defensively write your migrations as much as possible so that re-runs are relatively safe, though the tool helps to handle this.
One bit not mentioned, and particularly useful in more modern RDBMS with JSON binary expressions in the database are to leverage JSON columns and avoid joins altogether for a lot of use cases. There are a lot of times where you have variance of sub-information, or other data where table normalization and joins work against you. Even with indexes, joins are costly, especially under load at scale with millions of simultaneous users. You can avoid a lot of this by simply having that sub-table information inside a JSON field with the row in question.
For example, logs and notes related to a specific field. Variable transaction data (paypal vs amazon vs google payments), where the logs/details from the API aren't something that really needs to be in a separate table but related to the transaction.
Another would be something like a classifieds site where many fields are repeated, but sub-fields can vary dramatically by the type of item or category.
Knowing how/when to leverage denormalization and JSON can be one of the most impactful things you can do in terms of performance in practice, short of falling back to a search database (Elastic, Quickwit, etc), which can also be practical depending on your needs, but adds complexity.
Similarly, knowing how your datagase uses certain types of data/serialization... for example UUIDv7 if you don't mind storing creation time (utc) of a record, or COMB if using say MS-SQL in particular... the serialization of said field in practice helps in terms of understanding how indexes update and impact performance.
I do wish the guide was expanded a bit with lots of specific examples and details... a lot of it is hand-wavy blurbs.
In my experience, good monitoring is a must have from the very beginning. Eyball dashboard from time to time to spot issues, and use during incidents.
Things like connections stats, deadlock monitoring, slow queries. All come standard in AWS/GCP.
Interesting topic, We built and use DeepSQL (https://deepsql.ai/) at Stayflexi(YC) to address some of these issues. It's an DBA agent to prevent schema bloat, over indexing and does continuous monitoring of query workloads and proposes fixes.
Schema blot issues are real when you are vibe coding. Our engineers vibe coded and bloated our schema from 230 tables to 600+ tables. Many of them have repetitions of columns across the tables and often too much indexing. If your Postgres is on Aurora, the bloat easily multiplies your bills.
If you have a horizontally scaled app (many 100s of API servers and async workers) you’re also probably going to need a connection pooling proxy like pgbouncer! with separate pools for separate connection configs (lower/higher timeouts, reader/writer). There’s a section on this that’s a bit of a stub right now, but IME tuning and configuring these connection poolers is pretty nontrivial and worth an expanded section!
I’ve seen this pointed out in other comments but I’d also strongly recommend expanding with a section on monitoring and alerting. One could write a blog post almost of this length just on monitoring :)
Great blog! Thanks for writing this one up. Such a useful one for anyone who is build with Postgres. Succinctly reminds of all the battle scars working with many customers over the past decade. ;)
If your domain is analytics-heavy, don't try to optimize your Postgres for analytics. Follow the standard pattern of mirroring your data to a data warehouse and go to town there instead. There will be upfront cost of having to pay for a warehouse and the ETL, but it will be worth it.
The infrastructure decisions in early-stage startups are brutal.
Most founders I know underestimate Postgres connection pooling
until it bites them at 10x scale. PgBouncer saved us twice.This was a helpful guide. For someone using postgres for a few years, but rarely to its limits, a lot of it was review, but it had some great new tidbits to take in.
Lately I been questioning whether it’s actually a good idea to pool connections. Don’t your in the risk of leaking privileges or information from other requests?
> I’ve found normal forms to sometimes be at odds with query efficiency and ease of use, which is critical when you’re moving fast—sometimes it’s just easier to dump data into a jsonb column.
If you’re a startup, the performance cost of storing everything in JSONB is going to outstrip any gains you might get from denormalization. JOINs are simply not that hard if you design your schema intelligently. Additionally, allowing freeform text columns for things like statuses will eventually bite you with fun problems like `closed != CLOSED != Closed`.
> Use foreign keys with cascading deletes for low-volume tables, particularly where database consistency and correctness are important. Careful at higher volume.
Absolutely. Just be careful with 1:M, or M:N, for large values of M and N. You don’t want to trigger a surprise deletion of hundreds of thousands of rows.
> Indexes by default use a btree implementation. It’s most helpful to think of indexes as just another table in Postgres, with data stored in a specific format which is optimized for lookups (more on this later).
For a single row lookup (which is what this section was referring to), yes. For range scans, if the indexed column isn’t k-sortable, a sequential scan can start beating the performance of the multiple lookups pretty quickly.
> There are cases where you think an index should be used, but the query planner is still seq scanning anyway, despite table statistics being up to date and the index being valid.
This is usually caused by one of two things: forgetting that indices are (generally) B+trees and having data laid out in a manner that is inefficient for the query, or having data that isn’t uniformly distributed - for example, for some / many companies, the geographical distribution of users is going to be heavily clustered around more populous cities. Histograms are one way to deal with this.
Another topic not discussed in TFA is other index types - BRIN in particular can be incredibly performant while adding almost zero overhead, if the shape of your data makes sense for them (time-series is the obvious one, but anything with useful clustering should be considered).
All in all, this is one of the better tl;dr articles on Postgres I’ve read. Well done, Hatchet.
I'd say this applies to any database management system.
the first rule of database management is to not host or manage your database unless you are willing to pay someone to do it full time
Yea so like let’s spend on wars instead??
for people who have tried hatchet & restate - which one do you prefer ?
Migrating additional columns is interesting to avoid damaging the database.
miserable.
I would compare it to the number of lawyers over population
:(
TL;DR: use Oracle DB @ OCI Cloud, make yourself a favour later.
I did a search in that post for "function", zero results.
Unimpressive. Not even the most cursory of discussion of stored functions ?
Given that many startup's Postgres instances will no doubt be backing some web-ui or app that takes untrusted input, surely they could have at least had a brief discussion about how stored functions can help against SQL injection attacks ?
Not only that but it means you have to think, it prevents devs just writing their own random queries.
Also zero mention of `text`, which is highly encouraged in Postgres instead of the silly old `varchar(255)`