What MongoDB got right (2015)
blog.nelhage.com
blog.nelhage.com
They marketed their asses off while they iterated to build a document store that was worth the praise they got people to heap on to it.
IIRC the acquisition of the wired tiger engine was basically a rewrite, because the previous implementation was just bad. So if the database engine wasn't good, and distribution wasn't good (until later), and schema management wasn't good, and the query language wasn't all that great, what did mongo really get right? Marketing and being early to market/new and shiny at the right time (when everyone was deciding that writing the frontend and the backend in the same language, javascript, was an awesome thing -- which I don't necessarily disagree with).
If someone needs a database these days, I always just suggest postgres. if it's good enoguh for reddit (https://github.com/reddit/reddit/wiki/Architecture-Overview#...), it's good enough for the next facebook for dogs (unless your problem is completely different then of course you should explore whatever database paradigm fits your problem the best).
and if you have projects in maintenance using mongo, mysql. etc. try fdw[0].
It will allow you to query old systems using pgsql/plpython to integrate more easily into your new design or developments.
Most database systems came from relation theory background and optimize for performance, operations, data safety, but often developer convenience was sacrificed.
Did anybody liked writing data migration? Try explaining value of table migration to some Product Manager of early startup :-).
You can argue if that's were the right tradeoffs, but that's how MongoDB get a lot of steam.
I have never worked with a company that was sunk by the cost of migrating data between schemas. I have, however, worked with companies where the bugs that come with a schemaless database have almost sunk it. YMMV of course but I think it’s a false trade-off.
I would say that the developer experience was good in the sense that it was very easy to replicate Mongo databases - but that’s more of a system administration thing. I honestly don’t think that schemaless databases add any real value.
Mongo has its place, but its place isn't in most line-of-business applications for SMEs.
"Pick the right tool for the job" has been the biggest casualty in the midst of the schemaless hype we're going through right now.
Even getting to the point of realizing that most domains where NoSQL or schemaless solutions thrive are also domains where you need to be thinking in suites of persistence technology, ie where trying to choose a single data persistence technology is crazy, would be a big step.
A beast like Reddit (for example), has clear areas where a document centered schemaless approach would yield big wins, but that doesn't mean it's ideal for handling their advertising billing or their master data.
"The job" is quite often many smaller jobs, each of which needs it's own tool. Sometimes it's a lot easier to put on slippers than to carpet the whole world.
data integrity isn't good enough for a product manager? It's like saying "statically typed languages aren't worth the effort, do everything in JavaScript".
Furthermore, MongoDB or not, an application in development needs DB migrations. I don't believe one second you can change all your classes or structs and pretend like MongoDB will magically update your previous documents and collections to the current state of your application.
But one huge advantage of MongoDB is that it is schemaless. And so you have a lot more options for how you migrate documents. For example you can move all of the data under a particular key e.g. "v1" and allow two schemas to co-exist within the one document. Significantly less risky than running migration scripts or doing a alter schema in production.
What schemaless lets you do is hide your risk better by disguising those steps. "db.table.update({$set: ...}, {multi: true})" looks friendly, but it's the same terrifying operation as running DROP TABLE A the instant you're done generating table B. And it looks easier, but that's because your code's beliefs about which schema it's reading from are hidden and implicit.
That's absolutely true, but this is also what gave it a bad rep. It put security on the back-seat, and since it's "convenient" and just REST, it allows for moronic stuff like talking to your database straight from a client's web-browser. Security as an afterthought promotes bad or inexperienced developers to do silly stuff, and since MongoDB focuses primarily on web applications - that is a really really bad thing. How many leaks have there been simply because the database was exposed to the public internet without any authentication? Too many.
Now the problem is that inexperienced devs will flock to something that makes it that easy, and there's no real way of preventing that, but developers should never be put first - it leads to lazy solutions and compromises on too many other levels.
People are so spoiled this days. What is so hard in writing table migration, it is like a few minutes of your time. And why would some non techy product manager even know that you wrote some migration sql file. It is a minuscule technical detail related to the chosen technology and he will not see any repercussion in the production. But try to explain to him why after some reckless sprints you got many inconsistencies in the user facing app which you will have if you abuse "schemaless" NoSQL but use implicit schemas which change over time without proper checks. And if my product manager will push just for initial speed all the time without 0 time for some basic architecture (and I mean really basic), engineering practices and such, you know really useful stuff, I would leave that place without second thought. It is digging it`s own hole from the beginning.
And what is the problem of data migrations? Implicit schemas in MongoDB are much worse of a problem when your data requirements change. With relational databases you at least have integrity checks on the DB level that guard you from much of the possible problems when doing such migrations.
Wow, I wrote migration script for a couple of table rationalisation and it took me a week, to make it error-prone and to provide a revert option. Not to mention that you usually need to work with production data on your local to test if the migration will work without any errors.
And If you happen to have a migration error on your production database. It is a terrible experience, which usually includes downtime.
Maybe my lack of experience on writing SQL migrations increased that time, but isn't this the whole point of the OP in terms of dev first.
What about removing a column? What about if you've decided that you both current and next version of the code to work (then removing a column is a 3 step process, copy to new and use new, skip a version, remove old).
Migration can and is often hard.
Either you kick the can to the code, or do a migration script. Is there some magic style of coding that allows me to use the old data structures with the new code without having to do any special checks?
Well, there's your problem right there!
Even when you use a schemaless db like MongoDB, you still have to manage your data (let's say the schema is implicit).
If for example you decide that the "phone number" field is no longer a string but a list of objects (containing each the phone number and the kind of number), you will have to update the old data to adapt it to the new model, incurring in the risk of messing up production data.
Of course you can (and it is what many people do) avoid the problem at the data level and "fix it" at the application level (for example, checking every time if you have a single string or the new list), but if you really want you can do the same with a SQL database (keep the "phone" field on the table, but add the new table for the list of numbers).
A disadvantage of the SQL approach is that many are less versed in it than in the application language, but there are many tools available to handle SQL migrations in your framework's language.
Yes, the "secret" to MongoDB not having to write data migrations is simple: you get to simply have them as technical debt for later.
What could go wrong ?
Well, that's easy too: an accounting system where each method has a switch with 10+ cases, each case having 30+ lines of code inside of it, depending on something like "record version" or worse, on implicit tests.
And what could be worse ? Well, having a switch with only 9 cases in one of those methods ...
This is essentially how my team handles things while using HBase as our main DB. At first it felt weird to not write a migration at the DB level and maintaining compatibility is certainly something that you have to keep in mind at all times, but we have a pretty solid internal library for handling de/serialization in Scala.
> A disadvantage of the SQL approach is that many are less versed in it than in the application language, but there are many tools available to handle SQL migrations in your framework's language.
I was surprised by this at my previous job. One of the first things I learned about when starting to do web development was database migrations (Miguel Grinberg's Flask tutorial). Sure, I had to learn Alembic when I started the job, but it's a pretty straight forward API and the main issue was the lack of transactional DDL in MySQL that we constantly had to keep in mind. Besides that it was just writing SQL statements wrapped in Python.
I have worked on a number of large enterprise projects where there were hundreds and hundreds of migrations over the lifespan. I have never seen it take minutes of time. Or maybe you're just one of those fun developers who never properly tested their work.
But I have definitely seen production outages and subsequent failures to rollback because of edge cases between various individual migrations.
But often you are adding to reference data or running a transform SQL in order to produce some new denormalised table or the huge array of changes needed to support new use cases. These can't be automated and are hugely error prone.
Unless you migration involves "disable all checks, add new columns/tables, move and delete old data, enable all checks"
If you are migrating to a denormalised table then it should obviously crash when your conversion tool is doing something wrong and the RDBMS just prevented it from succeeding in doing something wrong.
If your migration is actually error prone as in "mangling and corrupting" data I suggest you are doing something wrong and you probably disabled the safety checks in your RDBMS.
And I have written many migrations in systems with tens of different db`s/schemas in the same project which could be different based on the geographical region or some other requirement and never had any big problem
Especially since you need to write the rollback script as well. And then test it with data. And then ensure there are no weird interactions with other migrations. Because if there is a problem with any one of them then it can be a huge headache to unwind everything.
...if you're making a relatively small change to a non-mission-critical, non-distributed system that can be taken offline briefly while you migrate; then you can just shut everything down, migrate, roll new versions, and start up again.
If you need to keep everything running while you do a gradual rollout across thousands of DBs, this gets more complicated very quickly - and that's not even taking into account more complex migrations that involve more than just "add / remove this column".
You're definitely right, however, to point out that implicit schemas don't make this better: you have all the same problems, but with less documentation / validation helping you not screw things up. It's also a good point that many businesses / applications never reach this latter stage where migrations become much, much more complex; when you can get away with the former approach, you definitely should.
The company I work for has 10's of TB's of data on most customer systems, and migration is either impossible as they have no place to store it, or too costly in time to move.
So while writing the code is not hard, the difficulty (or even possibility) of field upgrades can vastly influence whether or not a feature is considered.
Specifically, I would say mobile and node.js developers. The other part that worked well was so sharding, out-of-the-box. Oh, there were major issues like an index updates skipping documents, but it appeared to work and was the easiest to set up.
Painters sounds about right.
Whatever the inconvenience is, I need to be able to stand in for them being an acceptable price to pay for my personal convenience. And that is much more likely to be the case if I carry the "inconvenience" part of the equation, while others get the "convenience" part - like the operations people, or the end user, or other developers. That's because I consider myself a professional who's getting paid good money to make other people's life easier, especially if those "other people" vastly outnumber myself or the few other devs on my project team. That is the essence of my job, and it's not intended to be "convenient" for me - developer convenience is therefore one of the least-relevant criteria by which I ought to choose how I'm going to do it.
Developers who highly value "convenient" solutions are bloody amateurs in my book, especially if they can't even envision the ways in which their convenience has to be paid for by others.
Here's the thing, in practice relational databases aren't hard, nor will most devs ever need to know what relational theory is in order to be productive with databases built on it. Performance optimization does get a bit harder but there are relatively few situations where there's any magic involved, knowing how to read an explain plan and when/where to add an index is what most of it boils down to.
Yes, I like it. And you know what? My databases have lots of checks in the schema. This way I know, that I get correct data when I query the database. In MongoDB I'm not sure. I can throw there anything, and get anything. It's just a normal GIGO queue.
And the migrations are fine. You just need to know the tools. I observed huge programmers' aversion to learn SQL (however they don't complain when they need to learn HQL).
Disagreed. SQL is great for programmatically accessing data, especially with a good client library like SQLAlchemy. I have a much easier time understanding and, most importantly, debugging SQL than MongoDB's query language.
It did not solve the injection issue either, MongoDB injections are a thing[1] and they're much harder to reason about than SQL injections.
[1]: https://www.owasp.org/index.php/Testing_for_NoSQL_injection
SQL is just a language and needs to be learned. It's not hard and being declarative it also teaches to think differently than imperative programming.
Want to query in Javascript? Sure go ahead, but don't say SQL is harder.
Assembler is just a language and needs to be learned.
Oh wait. Nothing needs to be learned and there are always cases for higher up abstractions. Especially SQL which can get extremely complex if it needs to.
Some ORMs I've worked with allowed me to completely bypass the ORM system while still getting most of the benefits. I do recall a ORM that I've use a long while ago that didn't support reading Views as objects but it could be trivially added simply by writing a custom query method which was unpacked like any other query into the object and that worked for Views then.
The hard part of SQL is learning complex queries for "ad hoc exploration of data or for building aggregates over data sets": joins, unions, window functions, recursive common table expressions, etc. (Nevertheless it's still shorter than doing the same thing in some procedural language.)
The easy part is "programmatically accessing data." What is meant by "programmatically accessing data" anyway, besides "select * from . . ."? The writer's main criticism is:
> If you want to insert data,
> you’ll probably end up constructing the string (?,?,?,?,?,?,?,?)
> and counting arguments very carefully.
select *
from fruits
where color = :color and calories < :calories
MongoDB is excellent for design-first rapid agile development when domains are still maturing. This is common in large system redesigns and start-ups.
Comparing MongoDB to SQL or relational is pointless. They are tools. Given the right circumstances, both serve a purpose.
Relational databases are still necessary for analytics and reporting. (Try using any reporting tool on a schemaless database or a schema that’s a moving target)
I would argue that MongoDB is best used for R&D, but once your domains have matured, quickly moving to a relational data store will provide more sustainability.
I’ve also started incorporating graph databases into my architecture thinking. You should too.
Instead of using tables "Roles, Users, UserRoles, Actions, RoleActions" or whatever model you've come up with in the past, think in terms of graphs:
Person P [IS IN] Role R
Action A [IS IN] Role R
Now you can match a person to actions and using a query language like Cypher (Neo4J), it's very easy.
I've found this to be vastly more efficient than using a relational or document database.
I have an implementation of Identity Server 3 with the back end store in Neo4J and it's so much easier than managing a SQL back end.
You should try really hard to be unproductive with SQL and any relational system.
Uhm what?
At least in .NET+SQL Server land you can give your SQL queries named parameters. If there's a mismatch you'll get an exception saying so...
Perhaps this is a problem in some other stack (PHP?)
EDIT: To those mentioning migrations, note that reviewing migrations are like reviewing patches. Sometimes you want to not review a specific patch but rather look at the whole method, and `blame` a specific line or three to get an idea of what the previous developer's intention was. You can't get that type of overview from reviewing each patch or migration.
Add user-defined table types used as parameters and you've got a good way to pass batches of data too.
This is can be automatically handled if your ORM supports it; even a micro-ORM like Dapper does.
Otherwise my go-to solution for this if it isn't supported is to pass the collection as a user-defined table type, filled with the values. you can either use WHERE IN or join on this table variable.
EDIT: my perspective on this if from working with SQL Server and .NET. I don't know enough about Python or PHP development to comment on those.
You still have to write your SQL statements, and in your example you'd still be able to refer to your array by name:
select SomethingID where AnotherThing in @yourNamedArrayParameter
If you're not using Dapper just use a user-defined table type to hold your data and join on it instead of using WHERE IN.
I agree that manually building SQL strings can be cumbersome. But in practice you can use libraries that allow you to easily create them programmatically.
The JSON queries of MongoDB on the other hand, while they are simple to use programmatically, they are a pain to use outside of your program, for example for running ad-hoc queries, or generating a report from your data. And there is no library to fix this.
Suppose you built a "contact database," but all it was was the scanned images of people's business cards. You could flip through pages of them, but you can't do searches. But the input is so easy, just scan and save. And the output is likewise easy, just embed the images in a web page.
1. Running optical character recognition would be the first step in breaking down these cards into parsable information.
2. Converting the text into JSON and storing it in MongoDB would be taking it to the next level.
3. Breaking down the JSON into a set of tables would turn it into the most flexible and useful form --- but it's also the most work.
You don't have to take all your data all the way to tables. For example, I leave my website log files as text. Grep and awk are usually all I need. I don't usually import them into tables --- although I have a few times for easier analysis!
Although you don't have to turn everything into tables, I would say there are several applications out there where the developer should have gone all the way and not stopped at JSON.
Cassandra have this, any hash database should have some sort of unique concept or key, and IIRC graph database all node should be unique.
I didn't realize this was a rare thing in NoSQL world.
Nothing to do with the primary key and yes it is definitely rare in the NoSQL world.
That's only true if you're using PHP,Python,Javascript or dynamically typed language. Mongo documents don't automatically map to classes in C++ or Java, the impedance mismatch still exists and a data mapper library is still required, just like with RDBMS.