Thanks for writing this! The visualisations really drive a better understanding than pure text does, and it's quite clear that you have a better understanding of what database do under the hood than I do.
As such, I have a question for you: contrary to your article, I've always been taught that random primary keys are better than sequential ones. The reason for this, I was told, was to avoid "hotspots". I guess it only really applies once sharding comes into play, and perhaps also only if your primary key is your sharding key, but I think that's a pretty common setup.
I'm not really sure how to formulate a concrete question here, I guess I would like to hear your thoughts on any tradeoffs on sequential Vs random keys in sharded setups? Is there a case there random keys are valid, or have I been taught nonsense?
B+trees combined with sequential IDs are great for writes. This is because we are essentially just appending new rows to the "linked list" at the bottom level of the tree. We can also keep a high fill % if we know there isn't a lot of data churn.
If you're sharding based purely on sequential ID ranges, then yes this is a problem. Its better practice to shard based on a hash of your ID, so sequential id assignments turn into non-sequential shard keys, keeping things evenly distributed.
Thanks! Another great article! It strikes me that modulo sharding on a sequential id would probably work rather well, but it was not mentioned in this article. Is there a reason I'm not seeing that this is bad? I guess resharding might be problematic, as you can't easily split a shard in two without rewriting every shard if you do that...
yes, that's the crux of the problem. when you have a sharded database, typically you want to be able to add (and/or remove) shards easily and non-disruptively.
for example - your database is currently sharded across N nodes, and it's overloaded due to increased traffic, so you want to increase it to N+1 nodes (or N+M nodes, which can add complexity in some cases)
if adding a shard causes a significant increase in load on the database, that's usually a non-starter for a production workload, because at the time you want to do it, the database is already overloaded
you can read about this in the original Dynamo paper [0] from almost 20 years ago - consistent hashing is used to select 3 of the N nodes to host a given key. when node N+1 is added, it joins the cluster in such a way that it will "take over" hosting 1/Nth of the data, from each of the N nodes - meaning that a) the joining process places a relatively small load on each of those N nodes and b) once the node is fully joined, it reduces overall load evenly across all N nodes.
For our DBs (which are often unsharded), we've found the best performance using the user account ID as the first part of the cluster key and then a sequential id for whatever the record is as the second.
It's not as good as just a sequential ID at keeping the fragmentation and data movement down. However, it does ultimately lead to the best write performance for us because the user data ends up likely still appending to an empty page. It allows for more concurrent writes to the same table because they aren't all fighting over that end page.
Spanner in particular wants random primary keys. But there are sharded DBMSes that still use sequential PKs, like Citus. There are also some use cases for semi-sequential PKs like uuid7.
I remember this article for when I was researching for https://spacetimedb.com/. The interactivity is very cool, BTW!
One neat realization is that a database is in fact more about indexes than the actual raw tables (all things interesting work under this assumption), to the point that implementing the engine you get the impression that everything start with "CREATE INDEX" than "CREATE TABLE". This includes sequential scans, where as visualized in your article show that lay the data sequentially is in fact a form of index.
Now, I have the dream of make a engine more into this vision...
I've known for a long time that you usually want b-tree in Postgres/MySQL, but never understood too well how those actually work. This is the best explanation so far.
Also, for some reason there have been lots of HN articles incorrectly advising people to use uuid4 or v7 PKs with Postgres. Somehow this is the first time I've seen one say to just use serial.
> incorrectly advising people to use uuid4 or v7 PKs with Postgres
random UUIDs vs time-based UUIDs vs sequential integers has too many trade-offs and subtleties to call one of the options "incorrect" like you're doing here.
just as one example, any "just use serial everywhere" recommendation should mention the German tank problem [0] and its possible modern-day implications.
for example, if you're running a online shopping website, sequential order IDs means that anyone who places two orders is able to infer how many orders your website is processing over time. business people usually don't like leaking that information to competitors. telling them the technical justification of "it saves 8 bytes per order" is unlikely to sway them.
DB perf considerations aside, a lot of software pattern around idempotency/safe retries/horiz-scaling/distributed systems are super awkward with a serial pk because you don’t have any kind of unambiguous unique record identifier until after the DB write succeeds.
DB itself is “distributed” in that it’s running outside the services own memory in 99% of cases, in complex systems the actual DB write may be buried under multiple layers of service indirection across multiple hosts. Trying to design that correctly while also dealing with pre-write/post-write split on record id is a nightmare.
There are subtle gotchas around sequential UUID compared to serial depending on where you generate the UUIDs. You can kinda only get hard sequential guarantee if you are generating them at write time on DB host itself.
But, for both Serial & db-gen’d sequential UUID you can still encounter transaction commit order surprises. I think software relying on sequential records should use some mechanism other than Id/PK to determine it. I’ve personally encountered extremely subtle bugs related to transaction commit order and sequential Id assumptions multiple times.
I always thought this was too complicated to every really understand how it worked, especially the lock policy, but now with LLMs (assisted with sqlite’s very comprehensive comment policy) even a relative neophyte can start to understand how it all works together. Also the intro to the file is worth reading today:
* 2004 April 6
*
* The author disclaims copyright to this source code. In place of
* a legal notice, here is a blessing:
*
* May you do good and not evil.
* May you find forgiveness for yourself and forgive others.
* May you share freely, never taking more than you give.
*
*************************************
* This file implements an external (disk-based) database using BTrees.
* See the header comment on "btreeInt.h" for additional information.
* Including a description of file format and an overview of operation.
*/
"The deeper the tree, the slower it is to look up elements. Thus, we want shallow trees for our databases!"
With composite indices in InnoDB it's even more important to keep the tree streamlined and let it fan out according to data cardinality: https://news.ycombinator.com/item?id=34404641
I keep hearing about the downside of B(+)-Trees for DBs, that they have issues for certain scenarios, but I've never seen a simple, detailed list about them, what they are, and the scenarios they perform badly in.
It's really just a matter of tradeoffs. B-trees are great, but are better suited for high read % and medium/low write volume. In the opposite case, things like LSMs are typically better suited.
If you want a comprehensive resource, I'd recommend reading either Designing Data Intensive Applications (Kleppman) or Database Internals (Petrov). Both have chapters on B-trees and LSMs.
See my comment in the main thread for an example. In a worst case scenario, some data is simply too "frizzy" to index/search efficiently and with good performance in a B-tree.
I've read this paper and it's a neat idea. It hasn't been introduced into popular oss databases like postgres and mysql, and my understanding is it has some drawbacks for real prod use vs ths simplistic benchmarks presented in the paper.
Would love to know if anyones built something using it outside of academic testing.
As such, I have a question for you: contrary to your article, I've always been taught that random primary keys are better than sequential ones. The reason for this, I was told, was to avoid "hotspots". I guess it only really applies once sharding comes into play, and perhaps also only if your primary key is your sharding key, but I think that's a pretty common setup.
I'm not really sure how to formulate a concrete question here, I guess I would like to hear your thoughts on any tradeoffs on sequential Vs random keys in sharded setups? Is there a case there random keys are valid, or have I been taught nonsense?
If you're sharding based purely on sequential ID ranges, then yes this is a problem. Its better practice to shard based on a hash of your ID, so sequential id assignments turn into non-sequential shard keys, keeping things evenly distributed.
And since it's only used for speedy lookup we can even use a fast, cheap and non-secure hashing algorithm, so it's really a low-cost operation!
Thanks! This was really one of those aha-moments where I feel kinda stupid to not have thought of it myself!
https://planetscale.com/blog/database-sharding
yes, that's the crux of the problem. when you have a sharded database, typically you want to be able to add (and/or remove) shards easily and non-disruptively.
for example - your database is currently sharded across N nodes, and it's overloaded due to increased traffic, so you want to increase it to N+1 nodes (or N+M nodes, which can add complexity in some cases)
if adding a shard causes a significant increase in load on the database, that's usually a non-starter for a production workload, because at the time you want to do it, the database is already overloaded
you can read about this in the original Dynamo paper [0] from almost 20 years ago - consistent hashing is used to select 3 of the N nodes to host a given key. when node N+1 is added, it joins the cluster in such a way that it will "take over" hosting 1/Nth of the data, from each of the N nodes - meaning that a) the joining process places a relatively small load on each of those N nodes and b) once the node is fully joined, it reduces overall load evenly across all N nodes.
0: https://www.allthingsdistributed.com/2007/10/amazons_dynamo....
It's not as good as just a sequential ID at keeping the fragmentation and data movement down. However, it does ultimately lead to the best write performance for us because the user data ends up likely still appending to an empty page. It allows for more concurrent writes to the same table because they aren't all fighting over that end page.
UUIDv4 is madness.
One neat realization is that a database is in fact more about indexes than the actual raw tables (all things interesting work under this assumption), to the point that implementing the engine you get the impression that everything start with "CREATE INDEX" than "CREATE TABLE". This includes sequential scans, where as visualized in your article show that lay the data sequentially is in fact a form of index.
Now, I have the dream of make a engine more into this vision...
https://planetscale.com/learn/courses/mysql-for-developers
Also, for some reason there have been lots of HN articles incorrectly advising people to use uuid4 or v7 PKs with Postgres. Somehow this is the first time I've seen one say to just use serial.
random UUIDs vs time-based UUIDs vs sequential integers has too many trade-offs and subtleties to call one of the options "incorrect" like you're doing here.
just as one example, any "just use serial everywhere" recommendation should mention the German tank problem [0] and its possible modern-day implications.
for example, if you're running a online shopping website, sequential order IDs means that anyone who places two orders is able to infer how many orders your website is processing over time. business people usually don't like leaking that information to competitors. telling them the technical justification of "it saves 8 bytes per order" is unlikely to sway them.
0: https://en.wikipedia.org/wiki/German_tank_problem
DB itself is “distributed” in that it’s running outside the services own memory in 99% of cases, in complex systems the actual DB write may be buried under multiple layers of service indirection across multiple hosts. Trying to design that correctly while also dealing with pre-write/post-write split on record id is a nightmare.
But, for both Serial & db-gen’d sequential UUID you can still encounter transaction commit order surprises. I think software relying on sequential records should use some mechanism other than Id/PK to determine it. I’ve personally encountered extremely subtle bugs related to transaction commit order and sequential Id assumptions multiple times.
If it’s just being stored in the table, it doesn’t matter, but also if it doesn’t matter, just use v7.
Ideally you use IDENTITY with Postgres, but the end result is the same, yes.
https://github.com/sqlite/sqlite/blob/master/src/btree.c
I always thought this was too complicated to every really understand how it worked, especially the lock policy, but now with LLMs (assisted with sqlite’s very comprehensive comment policy) even a relative neophyte can start to understand how it all works together. Also the intro to the file is worth reading today:
* 2004 April 6 * * The author disclaims copyright to this source code. In place of * a legal notice, here is a blessing: * * May you do good and not evil. * May you find forgiveness for yourself and forgive others. * May you share freely, never taking more than you give. * ************************************* * This file implements an external (disk-based) database using BTrees. * See the header comment on "btreeInt.h" for additional information. * Including a description of file format and an overview of operation. */
With composite indices in InnoDB it's even more important to keep the tree streamlined and let it fan out according to data cardinality: https://news.ycombinator.com/item?id=34404641
But you'd rarely need it. We mostly have write intensive counters. We just write to redis first then aggregate and write to postgres.
This reduces number of writes we need in postgres a lot
If you want a comprehensive resource, I'd recommend reading either Designing Data Intensive Applications (Kleppman) or Database Internals (Petrov). Both have chapters on B-trees and LSMs.
Would love to know if anyones built something using it outside of academic testing.