[ACCUMULO-2519] FATE operation failed across upgrade

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 1.5.0, 1.5.1
• Fix Version/s: 1.5.2, 1.6.0
• Assignee: Sean Busbey
• Reporter: Keith Turner
• Created: 21/Mar/14 01:49
• Updated: 11/Oct/16 19:41
• Resolved: 05/Apr/14 01:00

Description

While running the new upgrade script I noticed that a FATE operation failed. I think this was caused by the package name changes in 1.6. However executing FATE ops across an upgrade is probably not safe, its certainly not tested or easy to test. Discussed this on IRC, should probably refuse to upgrade if FATE stack is not empty.

2014-03-20 18:20:40,724 [fate.Fate] ERROR: Thread "Repo runner 0" died java.lang.RuntimeException: java.lang.ClassNotFoundException: org.apache.accumulo.server.master.tableOps.TraceRepo
java.lang.RuntimeException: java.lang.RuntimeException: java.lang.ClassNotFoundException: org.apache.accumulo.server.master.tableOps.TraceRepo
        at org.apache.accumulo.fate.ZooStore.top(ZooStore.java:266)
        at org.apache.accumulo.fate.AgeOffStore.top(AgeOffStore.java:172)
        at org.apache.accumulo.fate.Fate$TransactionRunner.run(Fate.java:58)
        at org.apache.accumulo.fate.util.LoggingRunnable.run(LoggingRunnable.java:34)
        at java.lang.Thread.run(Thread.java:701)
Caused by: java.lang.RuntimeException: java.lang.ClassNotFoundException: org.apache.accumulo.server.master.tableOps.TraceRepo
        at org.apache.accumulo.fate.ZooStore.deserialize(ZooStore.java:79)
        at org.apache.accumulo.fate.ZooStore.top(ZooStore.java:262)
        ... 4 more
Caused by: java.lang.ClassNotFoundException: org.apache.accumulo.server.master.tableOps.TraceRepo
        at java.net.URLClassLoader$1.run(URLClassLoader.java:217)
        at java.security.AccessController.doPrivileged(Native Method)
        at java.net.URLClassLoader.findClass(URLClassLoader.java:205)
        at java.lang.ClassLoader.loadClass(ClassLoader.java:323)
        at org.apache.accumulo.start.classloader.AccumuloClassLoader$2.loadClass(AccumuloClassLoader.java:278)
        at java.lang.ClassLoader.loadClass(ClassLoader.java:268)
        at java.lang.Class.forName0(Native Method)
        at java.lang.Class.forName(Class.java:270)
        at java.io.ObjectInputStream.resolveClass(ObjectInputStream.java:624)
        at java.io.ObjectInputStream.readNonProxyDesc(ObjectInputStream.java:1611)
        at java.io.ObjectInputStream.readClassDesc(ObjectInputStream.java:1516)
        at java.io.ObjectInputStream.readOrdinaryObject(ObjectInputStream.java:1770)
        at java.io.ObjectInputStream.readObject0(ObjectInputStream.java:1349)
        at java.io.ObjectInputStream.readObject(ObjectInputStream.java:369)
        at org.apache.accumulo.fate.ZooStore.deserialize(ZooStore.java:77)
        ... 5 more

IRC converstation :

<busbey> hurm. so how useful would a test set that injects faults into the !METADATA table be?
<busbey> or into FATE
<busbey> for that matter
<busbey> to make sure that we have sufficient failure handling to avoid catastrophic loss
<kturner> I think I saw a FATE related bug in the logs also
<kturner> FATE serializes classes and pushes them on a stack in zookeeper
<kturner> in 1.6 package names were changed, so things could not deserialize
<busbey> oh boy
<busbey> that's not good
<busbey> so like they were serialized while the cluster was 1.5?
<busbey> and then post upgrade explosions?
<elserj> sounds like it
<busbey> were package names changed 1.4 -> 1.5 related to fate?
<kturner> yep
<busbey> because in theory
<busbey> I could have a 1.4 cluster
<elserj> almost want to preserve classes which were renamed as deprecated
<busbey> that I upgrade to 1.5 and then 1.6
<busbey> and I could, in theory not allow enough time for FATE to clear out in the mean
<busbey> well, or provide some kind of transition jar
<busbey> that includes classes to allow for burn off
<busbey> that you could later remove
<busbey> this sounds like a blocker
<busbey> barring some kind of documentation we could do
<busbey> for safely shutting down a cluster in prep for an upgrade
<busbey> the monitor doesn't show any indicators for waiting FATE operations, does it?
<kturner> no
<kturner> maybe 1.6 could refuse to upgrade if the FATE queue is not empty
<busbey> filed CCUMULO-2517
<busbey> well
<busbey> 1) was this also a problem doing 1.4 -> 1.5?
<busbey> and we just haven't had anyone hit it yet?
<elserj> do you have an idea of how many renames this introduces, keith?
<busbey> 2) that sounds like a good idea
<busbey> as a first check, then just say "please start up the master under PREV_VERSION" and wait for FATE to clear
<kturner> we could do the same thing for 1.5
<busbey> with a ref to upgrade notes that explain how to check if FATE is clear?
<kturner> yeah
<busbey> that will require we finish ACCUMULO-2469, I presume?
<busbey> (that's the ticket for documenting how to access zookeeper)
<busbey> two additional tickets or one?
<elserj> there's a class that will print fate ops
<busbey> 1) upgrade instructions should include how to check if there are fate operations pending
<busbey> 2) upgrade code should refuse to upgrade if there are fae operations pending
<busbey> nice! we could use that and leave 2469 for later, then?
<ctubbsii_bot> https://issues.apache.org/jira/browse/ACCUMULO-2469
<elserj> ctubbsii_bot you need to trim punctuation
* murraju (~[e-mail]) has joined #accumulo
<busbey> do those two sound like they cover the FATE bug?
<busbey> I presume we don't know enough yet to make a call on the delete marker thing?
<busbey> and that any additional guards on the GC should be aiming for post-1.6?
<kturner> I am creating a ticket, any problem w/ me just plopping this conversation onto the ticket?
<busbey> sounds good
<kturner> elserj?
<elserj> oh, sure

Issue Links

• breaks: ACCUMULO-4496 Upgrade fails due to outstanding Fate operations, refers to nonexistant explaination of how to handle
• supercedes: ACCUMULO-2140 Race conditions between client operations and upgrade
• links to: reviewboard