pycon 2012 what python can learn from java

Pycon 2012

What Python can learn from JavaJonathan Ellis / @spyced

(Not a web development perspective)

The power of tools

Remote debugging

✤ -Xdebug -Xrunjdwp:transport=dt_socket,server=y,suspend=n,address=1044

✤ 11 steps: “(6) install Wing IDE on the machine on which you plan to run your debug program... (8) copy wingdbstub.py into the same directory as your source files and import it in your Python source ... (10) In wingdbstub.py on your debug host, set kWingHostPort ... try setting kLogFile variable in wingdbstub.py for log additional diagnostic information.”

JMX

Triton

GC in a nutshell

✤ Mark✤ Sweep✤ Compact

Aside: reference counting

✤ Only good when allocation + assignment are relatively infrequent operations

In short

✤ Python has a terrible GC story✤ Poor instrumentation makes it worse

Troubleshooting OOM

Runtime profiling

✤ https://github.com/foursquare/heapaudit✤ https://github.com/mariusaeriksen/heapster✤ Commercial: AppDynamics, DynaTrace, others

Building blocks

✤ -javaagent✤ ASM: http://asm.ow2.org/

heapy http://guppy-pe.sourceforge.net/

Collections

✤ Less one-size-fits-all the more you care about performance✤ HashMap✤ ImmutableMap✤ ImmutableSortedMap✤ TreeMap✤ ConcurrentSkipListMap✤ ConcurrentLinkedHashMap✤ SnapTreeMap✤ NonBlockingHashMap

✤ Not to mention BiMap, Multimap, ...

Lists?

✤ ArrayList✤ CopyOnWriteArrayList✤ ArrayBlockingDeque✤ LinkedBlockingDeque✤ PriorityQueue, PriorityBlockingQueue✤ SynchronousQueue✤ TransferQueue (Java7)

Where are these in Python?

✤ Python developers want to write Python, not C✤ Similar question: “Why not Cassandra in Python?”✤ PyPy to the rescue?

✤ RPython extension methods

Growing a language (1998)

✤ Guy Steele: “I should not design a small language, and I should not design a large one. I need to design a language that can grow.”✤ http://www.cs.virginia.edu/~evans/cs655/readings/steele.pdf

✤ Currently, Python is not a growable language

Concurrency

✤ ExecutorService, ThreadPoolExecutor✤ FutureTask✤ Also: ScheduledThreadPoolExecutor

✤ ForkJoinPool

Concurrency

✤ For CPU-bound applications, copies are the enemy✤ Useful but dangerous:

✤ List.subList, NavigableMap.subMap✤ Corollary: you need to support threads + shared state

✤ Twisted✤ Actor model✤ Multi-process + sysv

✤ Local and remote computation: one size does not fit all

Copies are bad

✤ Copies of large things are especially bad✤ Remember fragmentation?

✤ Iterators (generators) are good✤ Java: ByteBuffer✤ Python: memoryview

The GIL

✤ The GIL offers negligible help writing concurrent code✤ Consider this trivial race condition:

if d[k] == v1:d[k] = v2

More subtle problems

✤ Which of these is threadsafe?✤ L.append(x)✤ x = x + 1✤ x += 1

More subtle problems

✤ Which of these is threadsafe?✤ L.append(x)✤ x = x + 1✤ x += 1

✤ “The thread safety of python operations depends on the compilation of python statements into byte-codes, which is an implementation detail and should not be relied upon.”

Are we stuck with explicit locks?

✤ ConcurrentMap✤ boolean replace(key, oldValue, newValue)

✤ NonBlockingHashMap✤ BlockingQueue✤ CopyOnWriteArrayList

Shared state good, Mutable state bad

✤ final: most under-appreciated language feature?✤ guava: Immutable collections

✤ http://code.google.com/p/guava-libraries/✤ Persistent collections

✤ (More accurately, “What Python can learn from Haskell”)✤ http://code.google.com/p/pcollections/✤ SnapTreeMap

Final thoughts

✤ JRuby is the most advanced Ruby implementation✤ Jython is less popular, but I’d rather write Java than C✤ Once you can write Python libraries that rival native ones

for speed, things will get much more interesting