Datomic Best Practices
docs.datomic.com
docs.datomic.com
I'm a big Rich Hickey fan. If you don't know who he is, he's the guy behind Clojure and Datomic. I don't use those tools, but his views on simplicity are wonderful.
Here's a great quote of his on the subject:
"Simplicity is hard work. But, there's a huge payoff. The person who has a genuinely simpler system - a system made out of genuinely simple parts, is going to be able to affect the greatest change with the least work. He's going to kick your ass. He's gonna spend more time simplifying things up front and in the long haul he's gonna wipe the plate with you because he'll have that ability to change things when you're struggling to push elephants around."
Here's his classic talk on simplicity if you haven't seen it yet: http://www.infoq.com/presentations/Simple-Made-Easy
It does bring an exceptionally elegant design (well worth reading Nikita Prokopov's "Unofficial guide" if you're curious). Also, the time and transaction-annotation features are unmatched AFAICT -- if you're working with complex data where provenance matters, Datomic can save a HUGE amount of work building tracking systems.
Neo4j, though marketed differently, is a similar approach (but the Community version is GPLv3 and Enterprise is AGPLv3). The Cypher query language is declarative in a similar way to Datomic - the biggest missing feature is transactions.
It's hard for me not to sympathize with him on this.
Sure, this was probably built up using simple experiments and designs. But consider the Mar's landing[1]. Simplicity would be to have a single mechanism for landing the Curiosity. Not 3. With one of them being a crane drop from a hovering rocket!?
I do feel there is an argument to up front simplicity. However, as systems grow, expect that the simplicity will be harder and harder to maintain and keep such requirements as performance met. To the point that it becomes a genuine tradeoff that has your standard cost/benefit analysis.
In the end, this falls to the trap of examples. If you are allowed to remove all assumptions from real use down to only a simple problem, you can get a simple solution. Add back in the realities of the problem, and the solution can get complex again. It is a shame that, in studies, so few real programs are actually looked at.
Why? Simple, in the way Rich Hickey advocates, means the opposite of complex, which means that things are woven together. You can have many landing strategies without them being tightly coupled together. A huge system isn't necessarily complex.
Mainly, the problem is that these speeches all talk about keeping things simple. In many problems, this can't be done. Understanding the simple helps. But the actual solution will not be simple. So any newspeak to get around that is just annoying.
I extend that into saying that people that can understand complicated things, as well, will have an advantage.
The idea to generate a new word that is hard to blur from existing ones and depends entirely on context is amusing in this context.
That is, what separates complicated from complex is one of context. Yet... contexts change. And often the first thing you do when building a solution to a problem is to reduce the problem to something easier to solve.
In this angle, I fully agree. Simplify your problem as much as you can. But do not be misled into thinking you can keep it simplified. As you add in more and more of the realities of the problem, they will reflect in the solution. And, often, the worst thing you can do is to try and cling to the "simple" solution that solved a different problem.
That is, understand the simple things well. See how they map onto the complicated things. Don't cling to the idea that they can be merely composed into the complicated solution. Often, several simple solutions can be subsumed by a more complicated one. Much in the same way that higher math can subsume lower maths.
The upsides:
SQL is a horrible language, yet all other noSQL DB also throw away the relational, transactional and ACID features that are great in postgres. Postgres with datalog syntax would basically be a win by itself. Datomic queries are data, not strings. Queries can be composed without string munging, and with clear understanding of what that will do to the query planner.
The schema has built-in support for has-one, has-many relationships, so there's no need for join tables.
I've never met a SQL query planner that didn't get in the way at some point. If needed, you can bypass the query planner, and get raw access to the data, and write your own query.
You can run an instance of it in-memory, which is fantastic for unit tests, so you don't have Postgres in production, but SQLite when testing.
The downsides:
It's closed source.
Operationally, it's unique. Because it uses immutable data everywhere, its indexing strategy is different. I don't have the experience of what it will do under high load.
The schema is 'weaker' than say, postgres. While you can specify "this column is type Int", you don't have the full power of Postgres constraints, so you can't declare 'column foo is required on all entities of this type', or "if foo is present, bar must not be present", etc. It should be possible to add that using a transactor library, but I don't think anyone has done serious work in that direction yet.
Compound indexing support isn't in the main DB yet. I had to write my own library: https://github.com/arohner/datomic-compound-index
The licensing is a huge pain in the ass. If I accidentally launch an extra peer over our license limit, our production environment will stop working until the extra peer comes down. This is really butting heads with the growing popularity of abstracting physical servers as clusters so I think the strategy is kind of a mistake on cognitect's behalf.
Don't run an SQL server as your KV store you'll likely screw up the config and performance will suffer. If you want competitive performance with other DB's you will likely end up running memcache between your KV and Query Engine(s).
Don't store data over 1KB. Yes, the database can technically handle them, but in real world applications and expected speeds it can't.
B-Tree Syncs can be slower then you think in surprising number of cases.
The "put your most restrictive clauses first" rule (which is reiterated on the new Best Practices page) usually seems to do the trick in our hands.
This particular thing seems like a backwards step though. One of the major reasons why relational databases became so popular was that: the user didn't have to think about which predicate to put first in the where clause, or which join to do first. Query optimizers can do that much better, and there is decades of research on how to optimizer relational queries (including Datalog queries -- see Deductive Databases). It is hard for users to make such decisions, and complex queries, views, runtime parameters make it near impossible to reason about performance of different queries.
I found working with Datomic really nice. I like how I can express queries using clojure. It also has some great performance for reads and the data cache cuts down on those reads from Dynamo, which keeps the AWS bill down.
I did not like the way schema worked in the beginning. You had to make sure you had it right from a very early stage of development and that can be difficult if requirements change, but they've addressed that in later versions. It's also probably the most complicated part of our infrastructure, which is 100% run on AWS. You've got 2 transactors (high availability) and X peers, you need to deploy a new version of Datomic on them. Coordinating that while minimizing downtime is no simple task.
So, he could patent any key technology, publish the implementation with copyright protection, give source/binaries to customers on condition they keep paying, let users extend it for internal use, and even let users submit such improvements for others to use.. His company continues to make money on the licensing in each of these cases. All of this has been done before. If anything, the real risk is on the users that the source license might change like what happen with QNX. It's why I advocate perpetual licenses for a given release at a given rate which are re-issued each year a client pays.
[1] https://www.schneier.com/blog/archives/2014/05/friday_squid_...