Initially motivated by the shutdown of the RethinkDB company, and the licensing situation with RethinkDB (a blocker for certain parts of my business), I worked very hard for two months to completely rewrite the realtime and database components of SageMathCloud (SMC) to use PostgreSQL instead of RethinkDB, initially motivated by this discussion in Hacker news. I battled with and used RethinkDB heavily since May 2015, and I’ve used PostgreSQL heavily as well, with production data, rewriting all the same queries in both systems, so I’m in a good position to compare them for my use case (the site SMC).
This is my story. It’s a personal comparison, with NO BENCHMARKS or hard data you could reproduce. It’s what I would tell you if we were talking by the water cooler.
- I’m very happy with the rewrite.
- Everything is an order of magnitude more efficient using PostgreSQL than it was with RethinkDB.
- It is much easier to do exploratory queries of our data using PostgreSQL than it was with RethinkDB. PostgreSQL is much more expressive than ReQL, has a massive number of built-in functions, so we are making much better use of our data. With RethinkDB, often we just ended up greping through the latest database dump.
- PostgreSQL is “statically typed”, whereas RethinkDB had no type or schema enforcement at all; explicit clear typing improved the quality and robustness of our application.
- We are saving $800 month (!), due to reduced CPU and disk space requirements.
- I had no clue that RethinkDB would be Apache licensed in February 2017.
A Mathematician’s Apology
This post is probably going to make some people involved with RethinkDB very angry at me:
“@williamstein for the best interest of rethinkdb community project, and especially if you respect the community and ex team members trying hard, please do not sway the community like this.” and “@williamstein can you delete your previous post…”
Not listening to users is perhaps not the best approach to building quality software. In Slava’s postmortem, he says:
“People wanted RethinkDB to be fast on workloads they actually tried, rather than “real world” workloads we suggested. For example, they’d write quick scripts to measure how long it takes to insert ten thousand documents without ever reading them back. MongoDB mastered these workloads brilliantly, while we fought the losing battle of educating the market.” - Slava
With SageMath we have had many very intense technical and other discussions with epic arguments back and forth. The one thing we don’t do is tell people not to even try to criticize our design choices: bring it on. Obviously, the Linux kernel is similar, and it is very successful.
I’m writing this blog post partly because I’ve said many positive things about RethinkDB. Really, what I love is the problems that RethinkDB solved, and where I believed RethinkDB could be 2-3 years from now if brilliant engineers like Daniel Mewes continued to work fulltime on the project. I don’t care so much that problems are solved using a particular piece of software. For example, every component of SMC has been rewritten multiple times – I’ve thrown away tens of thousands of lines of code. I care about solutions, not glorifying a particular piece of code for its own sake. Hence this post.
I can only hope that any devs who are really, really serious about RethinkDB having a future would listen to users, and hence will appreciate this post. But I’m also prepared to be hated for not staying silent.
Realtime web applications
There are many approaches to writing “realtime” web applications, i.e., event driven applications involving simultaneous multiple users, with the application updating quickly in response to what users do. After learning React.js and rewriting a lot of the frontend of SMC using React, I wanted to use a similar reactive architecture on the backend, where each component of the system listens for changes in state (the database), and reacts to it. In early 2015, I also wanted something like Facebook’s GraphQL, but there were no available implementation yet.
RethinkDB lets clients listen for changes in state to the database, and react to them. It was advertised as “production ready” in 2015, so I spent months rewriting SMC so it would use RethinkDB as the backend database. Before that, I was using Cassandra, and only making very simple use of the database, with all realtime functionality done at the application level in memory (so not using the database); that architecture worked OK, but there was a huge range of functionality I wanted to implement which was impossible to do with this approach, without introducing a message queue. Also, Cassandra was a bad fit for my data, and talking with the Datastax people on the phone about their pricing really scared me.
I spent the summer of 2015 rebuilding SMC on React + RethinkDB, with the RethinkDB rewrite being many, many months of hard work and debugging, basically from May 2015 to July 2016. I hit critical bugs that would crash RethinkDB in some edge case, which the RethinkDB devs would always fix. I also encountered a lot of painful scalability and performance issues, which I fixed by tedious benchmarking, debugging, studying logs, and introducing client side workarounds (e.g., idle timeouts on changefeeds). In July 2016, using RethinkDB become pretty stable.
Jonathan Lee, a computer science student working with me on SMC in Summer 2015, advised me against using RethinkDB due to performance issues. In particular, he pointed out this 2015 blog post, in which RethinkDB is consistently 5x-10x slower than MongoDB. I ignored Jonathan’s advice, because I believed RethinkDB would catch up within a year or two. I thought they would obsess over benchmarks now that they were production ready. I didn’t realize it would take nearly a year for them to fix the bugs in their automatic failover and stabilize the current features. I had a nagging feeling deep down that Jonathan was right and I was making a big mistake, but I ignored it.
A RethinkDB employee told me he thought I was their biggest user in terms of how hard I was pushing RethinkDB.
Using RethinkDB up until July 2016 was painful. I remember so many times doubling and doubling again the cpu’s in the RethinDB nodes, in order to handle the load from (say) 10K changefeeds. Maybe I just needed to be “educated” and was using RethinkDB incorrectly. Everything was a battle; even trying to do backups was really painful, and eventually we gave up on making proper full consistent backups (instead, backing up only the really important tables via complete JSON dumps). We also had a lot of issues with disk usage.
Around July 2016, I finally got a setup using RethinkDB to be stable and working. I finally learned to really appreciate Docker and Kubernetes, since they make it very easy to tweak dials to scale things up and down. Also Harald Schilly suggested using “RethinkDB proxy nodes”, inspiring this section of the RethinkDB docs. These are RethinkDB nodes that don’t store any data on disk, but do the hard work of processing and serving changefeeds:
“The proxy node can do some query processing itself, reducing CPU load on database servers.”
We ended up spinning up a Kubernetes cluster with 20 rethinkdb proxy/webserver pods, in addition to our 6-node RethinkDB cluster, and we could handle our load. Even then, the proxy nodes would often run at relatively high cpu usage. I never understood why. In fact, they were the only part of the entire SMC architecture whose high CPU usage I didn’t understand. By training and profession, I’m a pure mathematics researchers and lover of open source software, so I’m used to trying to understand how and why things work the way they do, but I never understood this.
When the RethinkDB company shut down, I initially decided to just wait and see what happened, maybe for a year or two, since our site was working fine with the many nodes mentioned above (I also didn’t realize how much money we were wasting on this setup). Then I was in a very long and intense meeting with a potentially major customer for an on-premises install, and one of their basic requirements was “no AGPL in the stack”. With the RethinkDB company gone, there was no way to satisfy that requirement, and my requests went nowhere at the time.
I had assumed that the speed would increase substantially due to focused work of Daniel Mewes during 2017. However, my understanding is that he went to work fulltime at Stripe, and will not be working on RethinkDB much. I also worried that the license situation wouldn’t be resolved: “Worrying about licensing is what PG would call a sitcom idea  – it feels like doing useful work, but in actuality it makes no difference whatsoever.”, though as we all know now it was just resolved!
So in early December 2016, I decided enough was enough, and I started rewriting our Rethink code, which is 5600 lines of CoffeeScript, to instead use PostgreSQL. I spent the first week making prototypes and benchmarks using the LISTEN/NOTIFY/TRIGGER functionality of PostgreSQL. For me, I realized the problem should not be “use some cool tech”, but instead “can I use this tech to solve my customer problems”. Even if LISTEN/NOTIFY/TRIGGER are much lower level, and take a lot more work and thought than RethinkDB changefeeds, I don’t care if the end result is better.
I learned from this discussion in Hacker news that PostgreSQL has some basic building blocks for implementing something like RethinkDB changefeeds. Searching online for uses of NOTIFY/LISTEN yields some relatively simple (but clear!) demos, which I was very thankful for. I did lots of benchmarks, and came to the conclusion that this could work.
I knew exactly what I needed to accomplish, since I had it all running on top of RethinkDB in production. So no design was really needed. The problem was clear. Do exactly the same thing, but using PostgreSQL’s LISTEN/NOTIFY and triggers instead.
Regarding PostgreSQL, I’ve used it off and on since the late 1990s (in fact, PostgreSQL started at Berkeley the same year I started graduate school there!). There have been steady but major improvements to PostgreSQL over the years, including very good JSON document support, replication, and clearly somebody spent work making their LISTEN/NOTIFY functionality fast. Thank you, whoever you are.
I didn’t seriously consider MySQL since it doesn’t have LISTEN/NOTIFY, and is also GPL licensed, whereas PostgreSQL has a very liberal license.
After running tests and studying the API, I estimated I could rewrite SMC on top of PostgreSQL in “one month of focused work”.
I made a plan and spent all December rewriting SMC on top of PostgreSQL. Indeed it took exactly a month of focused work to do the basic rewrite.
The design I used was to setup a small number of LISTEN/NOTIFY channels, which would listen for changes on a table, and send the primary key and optionally other small columns to each connected webserver. This meant that the total number of triggers and LISTEN/NOTIFY channels that the database manages is quite small – hundreds at most. When a webserver client gets a notification, it then decides whether it is interested in that record, and if so does a SELECT back to the database for the rest of the data, which it then sends out to clients. (The problem of deciding whether to do the further select currently involves an O(N) call of a bunch of functions that check equality; it could be done much more efficiently with a Bloom filter or hash table.)
In moments of frustration at the CPU usage of RethinkDB, I had imagined implementing something like the above on top of RethinkDB, but decided not to, since it is literally doing exactly what RethinkDB must be doing. When I started vaguely thinking through the details it seemed hard and complicated, and I was worried that it would be even less efficient than RethinkDB. Last August 2016, when I had dinner with Daniel Mewes, he surprised me by telling me that the RethinkDB proxy nodes were all receiving (and presumably doing something with) all of the data for all updates to all tables that had any changefeeds. Maybe this was why things were inefficient…
In any case, I wrote code that automates creation of all triggers to do listen/notify. I went through my tables, and made sure to implement enough changefeed-style functionality so that they did everything I need. Also, since I actually knew what I was building ahead of time (and was scared of having hard-to-debug problems in production), I wrote a large number of unit tests.
There was also a complicated “graph style” query of “all collaborators on projects”
that caused a lot of trouble with RethinkDB, often taking 10 seconds for certain users
with lots of projects (e.g., me with over 500 projects). It’s a query that is hard
to express efficiently, involving a join over two tables. Also, RethinkDB couldn’t
do a changefeed on that query, so when the projects that I user collaborated on changed,
I would have to kill the changefeed and recreate it. When rewriting everything,
I decided to just do things right if possible, and came up with a single data
structure that properly tracked all projects and collaborators of a given user
by just watching the whole accounts and projects tables, and properly updating
some data structures. The code is in
ProjectAndUserTracker here, and it works very well in practice. Obviously, again, this same code could have been
written on top of RethinkDB, and it would have helped a lot.
In any case, to build my application on PostgreSQL, many new small problems absolutely had to be solved, many taking a day of concentration. Rewriting all the code from scratch did clean it up a lot.
Also, I hope to build multimaster async replication on the above changefeed functionality. This will be important when SMC is geographically distributed. I also have plans to do a partial-multi-master async between the main public SMC and individual docker images that users run offline, which provide a genuine full offline mode, and also provide simultaneous editing of files (with multiple cursors etc.), but with all compute happening on the user’s local machine in a docker container, which has its own small local PostgreSQL instance. But that’s for 2018…
Next, in early January, I started the process of writing code to migrate all the data from RethinkDB to PostgreSQL, with minimal downtime (so one big migration, then incremental updates). I thought this would take a few hours, but it ended up taking nearly a month! I have a lot of data – one table had 150 million records in it… Another obstruction is that PostgreSQL is statically typed, whereas RethinkDB is very much not… and this exposed tons of subtle issues in my data. In addition, with PostgreSQL it was obvious and trivial to impose conditions on my data, e.g., all email addresses in the accounts table are unique, so of course I imposed constraints! – due to race conditions there were multiple accounts with the same email address in my RethinkDB data, so I had to write some (scary) code to deal with that. I also had to deal with things like null bytes in JSON strings, and timestamps in nested JSON data structures, and many other issues. I used a combination of relational columns and JSONB in some cases, which I’ll revisit later.
I did miss one critical subtle bug regarding timestamp precision in the PostgreSQL Node.js driver, which would cost me days of painful work to debug.
As I migrated my data from PostgreSQL, I found myself in a unique position. I had years of production real-world data in both RethinkDB and PostgreSQL. By this point, I knew both query languages pretty well. I did a lot of random queries of my data, sitting in both DB’s, and looked at the resulting times. PostgreSQL was faster, usually 5x faster, sometimes only 2x faster, and often even 10x faster. Definitely, the act of writing queries in SQL was much faster for me than writing ReQL, despite me having used ReQL seriusly for over a year. There’s something really natural and powerful about SQL. And, holy crap, PostgreSQL has a lot of built in functions that you can use in your queries… and you can add more via Python and many other languages (I haven’t done this yet, but I have dreams of hooking Sage into PostgreSQL).
I am not providing my data or one single proof of my claims about speed. Again, this is watercooler talk. I have a couple hours to share my experiences with the world, and then I have to get back to work.
Backups, which involve dumping full tables from the database, were an order of magnitude faster with PostgreSQL.
The total disk space usage was an order of magnitude less (800GB versus 80GB) – some of our tables had a lot of TEXT fields, and PostgreSQL automatically compresses those, which was a huge win. Also – to be fair, we had no redundancy with PostgreSQL, whereas 3x redundancy with RethinkDB. SSD disk space on GCE is expensive, so the reduction in disk usage is saving us a lot of money.
I (and the other SMC devs) run a lot of single-user RethinkDB databases for development purposes. PostgreSQL tends to use (at least) an order of magnitude less RAM to do the same thing.
In math software like Sage, I have seen these “order of magnitude differences in speed” with many implementations of algorithms over the years. Often the first Python implementation of an algorithm is nice and illustrative and works; then you re-implement it in Cython, change algorithms, etc., and end up with something that is 100x faster. This is just the normal experience I’ve had with math software. I imagine databases are similar. Using 10x more disk space means 10x more reading and writing to disk, and disk is (way more than) 10x slower than RAM…
I spent a huge amount of time worry about connection pooling with RethinkDB to get better concurrency, finally just writing my own. With PostgreSQL I don’t even bother, and instead each web server just has exactly one connection to PostgreSQL, and that is of course served by exactly one single-threaded process in the PostgreSQL server. The root problem is “make results fast for users”, not “have a lot of concurrent connections”. By optimizing everything, the load on the database and the web servers is now overall very low, and can easily be handled over a single connection. There simply is no need for a connection pool for my application, since PostgreSQL is so fast. It’s also actually really nice that one client web server can’t slow down the whole database.
Going live: things started to fail spectacularly
After all the awesome microbenchmarking above, I expected that when we went live it would be way more efficient than RethinkDB. On a nervous quiet Saturday morning, we switch the live production site over, and everything looked reasonably good for a while.
Then things started to fail spectacularly.
Every connection to the database was pegged at 100% cpu doing SELECT queries. I didn’t know what to do. It made no sense. I made the database server faster and spun up way more web servers, which basically worked… but seemed weird. I panicked for a while, mulled over the problem, and kept raising the number of web servers, etc. This sucked. I thought for a while the only solution would be to greatly reduce the number of SELECT’s in the changefeeds. Recall that changefeeds work by doing a SELECT to get more data when necessary.
After convincing myself not to give up and shut down SMC for good, I calmed down and studied a lot of logs and found a PostgreSQL query that was taking 15s sometimes and locking the other queries. It was a query involving a subquery; it finds all collaborators of a user - it’s exactly the one mentioned above that I couldn’t make a changefeed on with RethinkDB. I then tried an instance of this query directly in psql, and it took only a few milliseconds. Weird. OK, I tried it with some other parameters, and it suddenly took 15 seconds at 100% CPU, with PostgreSQL doing some linear scan through data. Using EXPLAIN I found that with full production data the query planner was doing something idiotic in some cases. I learned how to impact the query planner, and then this query went back to taking only a few milliseconds for any input. With this one change to influence the query planner (to actually always use an index I had properly made), things became dramatically faster. Basically the load on the database server went from 100% to well under 5%.
The Node.js PostgreSQL driver
The Node.js PostgreSQL driver claims the native bindings provide a “20-30% increase in parsing speed”. For my workload, especially reading BYTEA data (blobs), the speed increase is 600%. This was another observation I made by looking at log files.
With all these optimizations, the load on the web servers and database, even when we have 600+ simultaneous users, is barely anything!
All the code I wrote related to this blog post is – ironically – AGPL.
Basically it is everything that starts with
SageMath, Inc. owns all the copyright, so we could license under something else if somebody is serious about wanting to create a nodejs project on top of PostgreSQL to provide changefeeds. I’m too busy with my company to do that, but I would be supportive.
I have often said that “RethinkDB is the first database I ever loved”. In fact, it’s the reactive approach to databases based on changefeeds that I love (just like I love using React.js). I still very much love trying to solve this problem. If I were in charge of the RethinkDB project, I would delete much of the code and instead focus on the problem – changefeeds, and build solutions on top of PostgreSQL (and maybe other databases). I’m very thankful for the RethinkDB project for giving me the opportunity to spend time using this approach to DB’s, so I know how it feels.
Regarding automatic failover and multiple nodes, what really matters is that the site works for users. Google Compute Engine is so reliable that a single VM tends to stay up for hundreds of days (!), or if it goes down, it comes back very quickly. PostgreSQL also now has a very good Master/Slave story. It’s much more likely that of 6 nodes, something will go wrong with one of them, and though RethinkDB automatically fails over, it can take a while and leave clients in bad shape. Also, at our current rate of growth, and with current load, it’ll be a long time until one VM isn’t sufficient to serve everything; our workload is 90% read and 10% write, so PostgreSQL Master/slave would also very effective for us for scaling out.
In conclusion, I hope that this post tells you as much about SMC as it does about databases. Other take-aways: - focus more on the real problem. - prepare to throw lots of code away; writing the first version(s) is not wasted effort, it brings essential insight - once you know what the code will do, it’s a lot easier to write it in a way that supports testing and refactoring - open source is critical for solving deep problems - don’t be afraid to try alternative architecture