Slashdot Mirror
Oracle's Java Company Change Breaks Eclipse
crabel writes "In Java 1.6.0_21, the company field was changed from 'Sun Microsystems, Inc' to 'Oracle.' Apparently not the best idea, because some applications depend on that field to identify the virtual machine. All Eclipse versions since 3.3 (released 2007) until and including the recent Helios release (2010) have been reported to crash with an OutOfMemoryError due to this change. This is particularly funny since the update is deployed through automatic update and suddenly applications cease to work."
How is this Oracle's problem?
Poor planning. Eclipse should not use a 'company' field to be pulling key VM info from. And there should be another more particular way to acquire VM information applications require. That was a poorly thought out situation from the get-go, but Oracle was mightily short sighted for making this change without much testing of compatible apps. Mind you, it isn't their fault as such, but pissing off all of those using Eclipse is mightily retarded. While we're on the subject of retarded, automatic updates? You deserve what you get if you trust those. You should be damn sure an update is solid, stable, and won't give you a BOHICA experience before you apply it. No sympathy for auto-update users.... that's just bad planning as well. So: Oracle: Minor thumbs down. Eclipse devs: Thumbs up overall (except for bloating), but thumbs down for this one. Auto-update Users: Not bothering with a thumb, too busy ROFLMAO.
-- Mal: "Well they tell you: never hit a man with a closed fist. But it is, on occasion, hilarious."
Yes and no. While it's not the best practice to rely on some field assuming it'll forever remain static, if you read the bug report in TFA (surprise, surprise), you'll find this:
So, the reason they examine it in the first place is to know whether or not they need to set specific values that are supported by the Sun/Oracle JVM. It's not optimal, but I can't exactly fault them for that.
He who has no
To Oracle's credit, when Eclipse dev's reported the issue (http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=6969236) Oracle immediately reverted the change within 2 days (http://hg.openjdk.java.net/hsx/hsx17/baseline/annotate/1771222afd14/make/hotspot_distro). They could have argued that it was Eclipse's fault for depending on the value in the first place and that rebranding their VM is something they should be allowed to do. But they put the best interest of other applications first. Still, it raises an issue that no one has really bothered with before. There are many Hostpot "vendor specific" options that are very commonly used. Almost every large application would configure heap sizes. There should be a standardized mechanism to define these options and thus avoid these very problems.
One reason... security. Prevents a unstable application from growing out of control, causing the whole system to start paging which with a GC becomes a diaster, dragging the whole system to a hault makign it unresponsive. So you set a heap size to "more than you'll ever need" so that it aborts if something goes wrong. There are technical advantages too. But still... I agree. The fixed heap limits are more of a pain than a benifit, especially when the default setting for the client JVM was 64MB until recently because it handnt been changed since around 1997.
Didn't you hear, Java is the new COBOL
Someone in our company ran into this several weeks ago, and I had kind of a fun time tracking down the problem. The summary and most of the comments are missing a lot of details and nuance, which actually make this problem kind of interesting.
1) It wasn't even running out of memory
Sun/Oracle's VM implementation (HotSpot) has a concept of a permanent generation, which is separate from the rest of the heap and has its own maximum size. This generation holds stuff like the code cache and interned strings. Whether or not this is a good concept is debatable, and as far as I know, they are planning to do away with it in the future as JRockit and HotSpot merge. At any rate, this is the space that was filling up. This probably didn't happen very quickly on a normal Eclipse distribution, but with a lot of plugins installed (and thus a lot of classes being loaded) it crashed pretty quickly.
2) This is only because of somewhat subtle differences between the various VMs
HotSpot is the only major JVM I know of that has a PermGen space - J9 (IBM) and JRockit (Oracle, via BEA) don't have this concept. Thus the requirement to be able to behave differently based on which VM you are using. Being able to behave properly on multiple VMs is especially important for Eclipse because not only do they have a lot of people using it on HotSpot, but because it is the basis for IBM's RAD, they have a ton of people using it on J9 as well.
3) This problem is in the launcher, not Eclipse itself
So, the crux of the problem is that Eclipse needs to start a VM, and has to know the proper flags to pass to it *before* it starts up. A few people have suggested trying reflection or other runtime methods as a better way to solve this, but this ignores a) Once the VM has started up, you can't change the heap or PermGen sizes, and b) As far as I know, there is no way to query the VM at runtime to figure out what its underlying heap structure looks like - that is an implementation detail.
So, while it does kind of suck that Eclipse was relying on a vendor name, it is trickier to solve than it appears at first glance. The only really graceful ways I can think of to solve this problem rely on some changes to the VM spec.