Thursday, 20 September 2018

Java release chains - Splitting features from security

There is now a Java release every 6 months - March and September. It started with Java 9 and we're about to get Java 11. But should you jump on the release train? To answer that, we need to look at how Java's release chains are being split.

Looking back at Java 8

In the olden days life was simple. There was a "major" Java release every few years and it contained lots of new features, for example Java 5, 6, 7 and 8. Each major release included new JDK methods, new JDK classes, deprecations, new JVM features and new language features. However, life wasn't actually as simple as it seemed.

Looking at Java 8, once it was released there was a regular frequency of "update" releases. The most well-known of these were 8u20, 8u40 and 8u60. But there were also many others - 8u5, 8u11, 8u25, 8u31, 8u45, 8u51, 8u65, 8u66, 8u71, 8u73, 8u74, 8u77, etc. So, what was going on?

Well the plan was quite simple, just not that widely known. 8u20, 8u40, 8u60 and so on were "feature" releases, while all the rest were security patch releases. See the full table.

  • 8u20, 8u40, 8u60 and so on were "feature" releases - every six months
  • 8u5, 8u11, 8u25, 8u31 and so on were "security" releases - every three months, plus additional emergency releases

If you look closely, you can see a pattern. The first security release after a feature release had a number 5 greater (8u25 is 5 greater than 8u20). The second security release after a feature release had a number 11 greater (8u31 is 11 greater than 8u20). This left space for emergency security releases like 8u66.

So what was a Java 8 feature release?

Well a feature release was allowed to contain anything that didn't impact the Java SE specification. So, JVM or tool enhancements might be allowed, particularly if covered by a flag that was disabled by default. For example, the "endorsed-standards override mechanism and the extension mechanism" was deprecated in 8u40, 8u60 added a new IBM character set, and 8u181 removed the Derby database from the JDK bundle.

If you did notice these feature releases, it might have been because Java 8 lambdas were pretty buggy until 8u40 (in my experience). As such, I've often commented that 8u40 should have been named 8.1 (although logically 8u20 would have been 8.1 and 8u40 would have been 8.2). Such a version numbering scheme would have been easy to grasp for the whole Java ecosystem, but sadly didn't happen with Java 8.

Java 11 and the release train

The observant will notice that there was a feature release every six months and a security release every three months. But this is exactly like the "new" release train post Java 9!

The key difference between then and now is that from Java 9 onwards each six-monthly "feature" release can now contain any change. Thus major JVM, library and language changes can be introduced every six months, which wasn't allowed before. It also means there is a new Java SE specification for each feature release.

To handle this high frequency and still provide stability, every three years, which is every six releases, there will be an "LTS" release - long term support. As I've discussed before, the LTS name is rather confusing given that Oracle will not be providing security patches beyond six months on an "long-term" LTS release for $free. (Read the linked blog to understand how the OpenJDK will be performing that role from now on.)

Obviously with this new release schedule the question of naming arose. Firstly, according to Oracle we are not supposed to refer to any release as a "major" release. Instead, all should be considered to be "feature" releases, and of equal importance. This is emphasised by using a single incrementing number - 9, 10, 11, 12, 13 and so on.

But are all releases really of equal importance when some have long-term support (LTS)?

The answer is clearly no. For many large companies, the LTS releases are the only ones that they will consider adopting, either to manage churn or because they insist on purchasing a support contract for Java. Thus only one conclusion can be drawn:

Java 11 is a major release because it has LTS, not because of its feature set. Java 12, 13 and so on are less significant simply because they do not have LTS.

Now this doesn't mean that Java 12 is any less tested or is in some way unsuited for general use. Nor does it mean that Java 11 is bigger than any of the other versions. It is simply a statement of fact that because Java 11 has LTS (long-term support) it will be more significant, and see adoption by a different type of user.

Release chains

The support model is built on chains of releases. In order to pick up a security patch, you may also have to pickup other changes. The support chain for any Java release is simple enough. The higher the number, the more fixes/features it contains. Thus 8u20 contains everything from 8u5 and 8u11.

Using bold for major/feature releases and italic for security patches we can see the pattern:

  • 8, 8u5, 8u11, 8u20, 8u25, 8u31, 8u40, 8u45, 8u51, 8u60, 8u65, 8u71, ...

From Java 9 onwards there are effectively two separate patterns.

Firstly, there is the "release train" pattern if you want both features and security patches:

  • 9, 9.0.1, 9.0.4, 10, 10.0.1, 10.0.2, 11, 11.0.1, 11.0.2, 12, 12.0.1, 12.0.2, ...

Secondly, there is the "LTS" pattern if you only want security patches:

  • 11, 11.0.1, 11.0.2, 11.0.3, 11.0.4, 11.0.6, 11.0.7, ...

Oracle is focussing its $free efforts on the first "release train" pattern. After 11.0.2, it is ultimately the job of the OpenJDK community (probably led by Red Hat) to keep the chain in the second "LTS" pattern going.

Note the key difference though. The two chains effectively split features from security.

In Oracle's view, the old Java 8 chain and the new "release train" chain are essentially equivalent. I disagree, but I need another blog article to explain why (as this one is already too long!). Stay tuned.

Summary

There are now two separate chains of releases - the "release train" with both features and security patches, and the "LTS" with just security patches. Neither chain is equivalent to how it worked in Java 8, so you need to choose which chain to adopt carefully.

Comments welcome. For example, did an 8u feature release break your code?

2 comments:

  1. awesome job! perfectly solved all my doubts!

    ReplyDelete
  2. Very informative as always, thanks for taking the time to write this up Stephen.

    ReplyDelete

Please be aware that by commenting you provide consent to associate your selected profile with your comment. Long comments or those with excessive links may be deleted by Blogger (not me!). All spam will be deleted.