This is an unfortunately invasive change which converts all of our BlockId
strings into actual BlockId types. Here are some advantages of doing this now:
+ Type safety
+ Code clarity - it's now obvious what the key of a shuffle or rdd block is,
for instance. Additionally, appearing in tuple/map type signatures is a big
readability bonus. A Seq[(String, BlockStatus)] is not very clear.
Further, we can now use more Scala features, like matching on BlockId types.
+ Explicit usage - we can now formally tell where various BlockIds are being used
(without doing string searches); this makes updating current BlockIds a much
clearer process, and compiler-supported.
(I'm looking at you, shuffle file consolidation.)
+ It will only get harder to make this change as time goes on.
Since this touches a lot of files, it'd be best to either get this patch
in quickly or throw it on the ground to avoid too many secondary merge conflicts.
Add an optional closure parameter to HadoopRDD instantiation to use when creating local JobConfs.
Having HadoopRDD accept this optional closure eliminates the need for the HadoopFileRDD added earlier. It makes the HadoopRDD more general, in that the caller can specify any JobConf initialization flow.
Address review comments, move to incubator spark
Also includes a small fix to speculative execution.
<edit> Continued from https://github.com/mesos/spark/pull/914 </edit>
Standalone Scheduler fault tolerance using ZooKeeper
This patch implements full distributed fault tolerance for standalone scheduler Masters.
There is only one master Leader at a time, which is actively serving scheduling
requests. If this Leader crashes, another master will eventually be elected, reconstruct
the state from the first Master, and continue serving scheduling requests.
Leader election is performed using the ZooKeeper leader election pattern. We try to minimize
the use of ZooKeeper and the assumptions about ZooKeeper's behavior, so there is a layer of
retries and session monitoring on top of the ZooKeeper client.
Master failover follows directly from the single-node Master recovery via the file
system (patch d5a96fe), save that the Master state is stored in ZooKeeper instead.
Configuration:
By default, no recovery mechanism is enabled (spark.deploy.recoveryMode = NONE).
By setting spark.deploy.recoveryMode to ZOOKEEPER and setting spark.deploy.zookeeper.url
to an appropriate ZooKeeper URL, ZooKeeper recovery mode is enabled.
By setting spark.deploy.recoveryMode to FILESYSTEM and setting spark.deploy.recoveryDirectory
to an appropriate directory accessible by the Master, we will keep the behavior of from d5a96fe.
Additionally, places where a Master could be specificied by a spark:// url can now take
comma-delimited lists to specify backup masters. Note that this is only used for registration
of NEW Workers and application Clients. Once a Worker or Client has registered with the
Master Leader, it is "in the system" and will never need to register again.
A fast and low-memory append-only map for shuffle operations
This is a continuation of the old repo's pull request https://github.com/mesos/spark/pull/823 to add a more efficient hashmap class for shuffles. I've optimized and tested this more thoroughly now so I think it's good to go. I've also addressed some of the comments that were outstanding there.
The idea is to reduce the cost of shuffles by taking advantage of the properties their hashmaps need. In particular, the hashmaps there are append-only, and a common operation is updating a key's value based on the old value. The included AppendOnlyMap class uses open hashing to use less space than Java's (by not having a linked list per bucket), does not support deletes, and has a changeValue operation to update a key in place without following the hash chain twice. In micro-benchmarks against java.util.HashMap and scala.collection.mutable.HashMap, this is 20-30% smaller and 10-40% faster depending on the number and type of keys. It's also noticeably faster than fastutil's Object2ObjectOpenHashMap.
I've also tested this in Spark apps now. While the speed gain is modest (partly due to other overheads, like serialization), there is some, and I think the lower memory usage is worth it. Here's one example where the speedup is most noticeable, in spark-shell on local mode:
```
scala> val nums = sc.parallelize(1 to 8).flatMap(x => (1 to 5e6.toInt)).cache
scala> nums.count
scala> def time(x: => Unit) = { val now = System.currentTimeMillis; x; System.currentTimeMillis - now }
scala> (1 to 8).map(_ => time(nums.map(x => (x % 100000, x)).reduceByKey(_ + _).count) / 1000.0)
```
This prints the following times before and after this change:
```
Before: Vector(4.368, 2.635, 2.549, 2.522, 2.233, 2.222, 2.214, 2.195)
After: Vector(3.588, 1.741, 1.706, 1.648, 1.777, 1.81, 1.776, 1.731)
```
I've also run the spark-perf suite, enhanced with some tests that use Ints (https://github.com/amplab/spark-perf/pull/9), and it shows some speedup on those, but less on the string ones (presumably due to existing overhead): https://gist.github.com/mateiz/6897121.
