Hacker News Reader

Comment by ante_annum | Hacker News Reader

So the explanation is:

  Let me be blunt -- sun.misc.Unsafe must die in a fire.
  It is -- wait for it -- Unsafe. It must go. Ignore any kind of theoretical rope and 
  start the path to righteousness

And what the author takes away is:

  This engineer hates the Unsafe class for no reason at all.

The reason the engineer hates it is that it's unsafe. Disagree with it, sure. But it's a reason, and it's a good reason (imho). I understand that it's commonly used by apps I use, but it's hard to keep reading after this kind of knee-jerk.

View on HN

In addition, it's in the sun.misc package, if it's not in java.* or javax.* it is ripe for pruning, you'd think people would have learned by now to stop using internal classes since pretty much every JRE/JDK release known to mankind has changed them or removed them.

It's sort of amazing that the diatribe here is against Oracle for wanting to remove an undocumented internal API because of typesafety, and not against the open source projects that use a specifically unsafe, undocumented internal API. I mean, nothing against these projects -- I understand that type safety can carry a performance penalty -- but the anger of this article is amazing.

The issues is 2 fold for people.

1) People use Apache Storm and Cassandra daily. These need Unsafe to function at current performance changing its functionality may change these tools. This can be considered existiential threat to their job.

2) The one big selling point of the JVM is that JVM 1.0 code will run in JVM 8.0 without issue, and often much faster. With this change that isn't necessarily true, and potientally when your .jar was compile will weigh on which JVM is used or HOW that jvm is started.

1. Have the maintainers of Storm, Cassandra, etc. said whether they architecturally need Unsafe in its current form, and whether the replacement APIs under development will fulfill their needs?

There's a lot of software out there that uses MD5 or DES right now. Saying that the software needs MD5 or DES is a different claim entirely.

>Have the maintainers of Storm, Cassandra, etc. said whether they architecturally need Unsafe in its current form [...] ?

As there is no alternative to unsafe serialization in terms of performance. It goes without saying.

>There's a lot of software out there that uses MD5 or DES right now. Saying that the software needs MD5 or DES is a different claim entirely.

I agree, but you are putting the trees before the forest.

Saying implementation X is flawed is one thing. Saying We are removing implementation X and offering no alternatives is another. While there are plan to offer an alternative, they are simply plans not an immidate alternative, thus a problem. When DES/MD5 were removed it was clear that they were already inferior implementations, and superior alternatives existed. That isn't true for java.

> As there is no alternative to unsafe serialization in terms of performance. It goes without saying.

I don't think that's true, though perhaps I don't understand what you mean by "unsafe serialization".

First, I suspect one can use JNI and get at least as good performance. There are other tradeoffs, like the ease of packaging, but in order to evaluate those we need a clear problem statement.

Second, premature optimization is the root of all evil, and we should be looking at realistic benchmarks to determine that there is a performance gain. There are many documented examples of software safety checks adding negligible performance overhead because they get branch-predicted away. Alternatively, new hardware support like Intel's MPX adds bounds-checking at the instruction level, so if we're talking about raw writes to a bounds-checked buffer (which is a vast improvement over raw writes to anywhere), this can probably be implemented efficiently on new hardware. On old hardware, the JVM could implement these as unchecked writes, which preserves the exact same performance, but the API would now have information to add safety where it is performant.

Finally, sun.misc.Unsafe covers a lot of ground. If we can restrict it to the specific uses that these projects make of it, that's still an improvement. I don't think these libraries use the whole thing, do they?

Very few things in science go without saying, and software performance and correctness is science.

Seeing as you've admitted you don't understand what I'm talking about. Here is some reading material:

http://www.javacodegeeks.com/2010/07/java-best-practices-hig...

Thanks, I definitely understand this topic imperfectly -- I am familiar with high-performance serialization, but I'm not familiar with Java or best practices with doing this sort of thing in Java. So it's definitely helpful for you to be clear in what you're discussing and what Unsafe is currently used for.

Is the specific Unsafe use here to serialize a Java object into a byte array? The linked article doesn't describe using Unsafe, so I'm still not sure what functions are at issue.

This is what I was referring to about a bounds-checked buffer in my parent comment: it seems like you could state that accesses to an area of memory are unsafe and unchecked, except to check that the reads/writes are within the buffer. This preserves safety for the JVM as a whole, but gets you native performance within the buffer. Intel MPX should be able to implement this efficiently, and the standard techniques in other languages for efficient bounds-checked memory access should all apply.

Am I completely off-base here?

You are. Here's a great idea, if you've never implemented a JVM don't tell somebody how it works.

Java VM isn't native code. Serialized objects are "Field"(s) within the JVM. Which is short hand for a "Growable Page File" more or less. Which starts at a fixed 32bit address. How does that work? JVM implementation takes care of that, so you decide.

On top of that whats inside the Field itself doesn't actually matter to anyone provided the same read/write opcodes do the right thing(s).

The serialization process starts by jumping to your initial field, copying the serialized data into another buffer. While checking integrity, and assuring that your data conforms to the serialization standard b/c its in memory representation doesn't. The fun part is when ever it encounters a RetAddr type it has to jump to that Field, and serialize THAT Field also, as that Field is part of the object your serializing. Of course you don't actually know whats in the Field without consulting its ClassField which gives you a basic prototype of what-is-where.

Furthermore what ever optimizations the JIT makes, it'll end up treating Fields more-or-less like stack frames. So you have to keep an active record of how it'll mangle memory, so you know how to untangle that when you want to do serialization.

Understand now?

About #2: there are many internal changes even between minor releases and if you relying on those internals, your code is NOT portable.

sun.misc.Unsafe effectively became a non-internal API a long time ago. Yeah, it's flagged internal, but it doesn't get broken without a damned good reason (I can't remember the last time it meaningfully broke, even).

1. it's quite likely that the Unsafe classes will be accessible with a flag, so for this kind of projects it's not really a big problem, as they provide the shell scripts for launching the jvm with all the necessary parameters.

On #2 - Java 1 code using the public API (java.*) will run in Java 8. Using internal APIs was never considered forward compatible.

The problem with that, is unlike most of the other internal classes, there is zero other way to do the things people are doing with unsafe without it.

It's not a matter of expedience, it's a matter of necessity.

Yes, removing Unsafe would make things like PowerMock and Robolectric impossible to (re)implement. So what? Those libraries never should have been possible, and we can live without them.

Languages need to be selective about what features they support, and Java was never intended to support things like mocking static methods, or creating instances without invoking any constructors.

Without the current unsafe package most performance sensitive applications (whether throughput, latency, or pause sensitive) will have no recourse to meet their current performance levels.

That will leave people to choose between staying on Java 8, worse (and often unacceptable) performance, and migrating to a different language.

At this point, there is a very good chance if your library/application is known for performance and exists on the JVM it uses unsafe and without a migration plan all of us who care about performance will likely have to move off the language.

With some effort, I'm sure the folks behind Cassandra, Hadoop etc. could achieve acceptable performance with allocateDirect(). It doesn't support explicit freeing, but you can always reuse buffers and write your code to allocate slices within them.

We can live without Robolectric? How do we write tests for Android code without it?

Turn on unitTests.returnDefaultValues, write some mockable wrappers around core classes like SystemClock, unit test what you can, and write instrumentation tests for everything else.

An increasing number of teams are moving away from Robolectric anyway, since it's so complex and buggy, and there are some things it will never properly support such as COLLATE LOCALIZED.

That is a pretty crappy explanation, on the engineer's side.

To be clear, any explanation that is not immediately followed with examples to how current valid uses of Unsafe can be accomplished without it, are simply witch hunts by an engineer that hasn't had to meet the same requirements.

It's a) unsafe, and b) an internal, undocumented API.

When you start importing from sun.misc.* you've made the conscious (or ignorant) decision to write non-portable code which may well break due to any JVM release. It sucks that these people are going to have to update their code (or that end users won't be able to upgrade to Java 9 without adding special command-line flags), but it's not due to fickle Oracle developers, it's due to technical debt that's finally catching up to them.

You quite simply have NO option to do on the JVM most of the things people are using unsafe for in Java 9/10 without a migration plan from Oracle.

What Oracle is close to "hey all you performance sensitive folks, we aren't going to support unsafe anymore, can you tell us what you are using it for and how we can migrate to that explicitly?"

It just so happens that one of the leading proposals is "leave in java 9 but make it explicit to turn on". This is probably the right way to do this (in my opinion) but it does have the downside of making neither camp happy (it leaves unsafe so the purists are mad, and it causes headaches for people using unsafe).

Whats wrong with:

- deprecate in Java 9

- "make it explicit to turn on" in 10, offer alternative official APIs as fallback

- remove entirely when most people have switched

Nothing at all. That's what I assume will happen at the end of the day. But people saying that we can just turn it off in 9 because you shouldn't be using internal api's don't seem to get that there is no external solution.

Why should third-party libraries pay for Oracle's technical debt?

It's not Oracle's technical debt, it's the developers of the third-party libraries that are importing from sun.misc.*.

I think the author's point about being able to make safe Unsafe replacements and keeping Unsafe around (for the time being) to enable better compatibility was thoughtful.

Reply on news.ycombinator.com