For me the bottom line of this video is: Microservices come with a huge amount of challenges, that you do not have with a monolith. Chris covers a lot of topics, each of which deserve a number of dedicated videos.
Most people have probably come across what is usually called a UUID (universally unique ID) while using software. UUIDs are typically a cryptic combination of alphanumeric characters and do not make any sense to the human brain. But why are they such a critical aspect to most computer programs?
Their purpose is pretty obvious: be able to identify a set of data (money transfer, customer, product, order, etc.) on a low technical level. The human brain, for most scenarios, does not need such an artificial construct but works nicely with the underlying “real” data. We identify a customer by looking at first name and surname. And if we have multiple customers “Mike Smith”, we add the date of birth. If that is still not enough, then there is the current address. And so on.
For the purpose of this discussion a customer’s UUIDs is not to be mixed up with the customer number but exists in addition. This may seem like overhead, but think about what happens when an organization buys a competitor. With a bit of “luck” there will be overlap between the customer numbers. Without a UUID already in place, all sorts of ugly workarounds need to be implemented under great time pressure, to be able to merge the customer lists then. If that happens, there is considerable risk of something going wrong, resulting in the loss of customers.
It would of course be possible to replicate the human brain’s approach of looking at data in their individual context. But that would make things unnecessarily complex, plus require a different approach for each type of data. So we help ourselves with a technical ID that is guaranteed to be unique. Generating such an ID is surprisingly complex, once you realize what the algorithm needs to accomplish:
- Be fast: There are many scenarios where you need to create tens of thousands of UUIDs per second (e.g. high-frequency trading, payments processing, telco billing, etc.). But “randomness” usually requires the use of cryptographic functions, which are notoriously expensive operations. In recent years this has become less of a concern, though, since many CPU now offer dedicated support here.
- Be unique across all computers that are involved with the application: While it is probably rarely a problem if two identical IDs are issued for two completely disparate organizations (ignoring scenarios like EDI), there are many cases where it is still highly relevant. Most critical applications run on more than one computer for high-availability and load-balancing purposes. So obviously there must never be a case where IDs clash. Also, it would likely cause problems if the same ID existed not only on the production system but also on a development or test system.
- Be relatively short: Many UUIDs are between 30 and 40 characters long, which is really not long, given that it is guaranteed there will never be a clash.
Let’s now look into the use of UUIDs. Apart from pretty obvious things like the aforementioned customer etc., they are used in very many systems for internal purposes. A good example are relational database managements systems, where each record (aka row) has its own ID. The same is true for messaging system (think JMS or MQTT).
The two core use-cases I see for those internal IDs are fault diagnosis and linking data. In today’s world most systems are highly distributed, even without the use of a micro-services architecture (which increases the level of distribution by orders of magnitude). To track a business transaction across multiple systems, you need to be able to identify these sub-transactions and the means for this are UUIDs. Ideally you have an operations console that automatically connects things between systems. In reality, though, there is often a lot of manual work to be done.
Another example of linking data together is master data management (MDM). Many organizations have done something in that area and most have failed. The core reason in my view is the approach. It is a business problem that is very closely linked with many technical challenges. And most organizations are bad at dealing with such a combination. There are more aspects, but I will cover those in a separate article.
Back to UUIDs. It might be tempting to leverage internal IDs (e.g. from a database system) for your application. But be warned, this is a very dangerous road. Those IDs are guaranteed to be unique only in the context, for which they are created, but not outside. Even more critical is using just a part of the IDs, because the rest seems to be a fixed value. I have seen a business-critical end-user application where part of the database’s row ID (Oracle Database v7) was used. Later the database was migrated to a higher version (Oracle Database v8) where the UUID algorithm had be changed. So the sub-string of the row ID was suddenly not unique anymore. The end-user application did not expect duplicates and crashed immediately after starting.
While at the subject of databases, there are people who like to use sequences as UUIDs. Sequences are numbers, which the database auto-increments and they seem a convenient and efficient way to obtain a unique ID. But there are various problems with that approach. Firstly, the ID is only unique within a single database instance. This typically creates all sorts of problems for testing the code, and also when moving it to production. Secondly, this kind of feature, while available in many database systems, is a proprietary extension of SQL. So you create yourself unnecessary problems for using different systems. Many organizations have standardized on one database system for production use. Having to use this also for DEV, CI, SIT, UAT, etc. may make things more difficult than necessary. More importantly, though, it increases the vendor lock-in with all the associated issues.
Let me finish with timestamps. They are the original sin of UUIDs. Really. People like them because they are human-readable, allow easy sorting of transaction into the order of processing, and just seem to be THE obvious way to go. But they are not unique! If your development machine is slow enough, relative to the transaction’s processing time, you may indeed not have issues. But that is only because at least one millisecond (you don’t use a resolution of seconds, do you?) goes by between transactions. A production machine, however, will likely be much faster. And what if multiple machines are working in parallel?
In one case I have seen there was considerable data loss, because someone had been clever enough to use a timestamp with a resolution of only seconds as the filename for writing PDFs into a directory. From there an archiving solution then picked them up for storage to fulfill a legal requirement. This guy’s notebook had been slow enough (it was in the early 2000s) that all files had been several seconds “apart”. But the production machine was a beefy server and it took several weeks until someone realized what had happened. Tens of thousands of documents were lost forever.
I hope this quick overview provided some value to you and will help you in the next discussion on why you really need a proper UUID.
While watching the video below, I was intrigued, when this interview questions was discussed (timestamp 56:30-57:35).
It made me think about my response, had I ever been asked the same. And it did not take too long before the answer was clear: the B-Tree. There is a very good section on it in Martin Kleppmann‘s book “Designing Data-Intensive Applications“. And I highly recommend this book anyway.
But as a starting point on B-Trees the following video is also quite helpful:
If you want to understand how graph databases might be able to help you…
Here is another interesting presentation from Martin given at the Code Mesh Conference 2015 in London.
Here is a rather interesting video from Martin Kleppmann where he talks about dealing with concurrent changes to data. While the title may sound theoretical to some, it is a topic that probably every developer has come across. And here is also the link to the paper with the algorithm presented. If you are interested in an implementation, check this Github project.
Sometimes people wonder why Oracle DB (both 10g and 11g) will not start after installation on Linux/Unix. In many cases the simple reason is the content of /etc/oratab. For each database it contains a line in the format
$ORACLE_SID:$ORACLE_HOME:START_DB_FLAG:The last column is set to “N” by default. Just change it to “Y” and run $ORACLE_HOME/bin/dbstart again. Your database should come up now.