A question about thrift performance.

[Vincent]

Hi, all

I have issue of thrift-performance in python, does anyone has an experience on thrift-in-python?

My question in stackoverflow:

http://stackoverflow.com/questions/14171227/why-is-thrift-binary-protocol-serialization-so-much-slow


Copy the question to here(open stackoverflow to check pretty-print question content):

I'm newbie on thrift. I wrote a thrift server in python, also client in python too.

Here is my thrift defination:
struct RatingByReport {
1: required string ticker,
2: required i32 cnt_institution,
3: optional list<string> strong_buy,
4: optional list<string> buy,
5: optional list<string> neutral,
6: optional list<string> sell,
7: optional list<string> strong_sell,
8: optional i32 cnt_maintain,
9: optional i32 cnt_upgrade,
10: optional i32 cnt_downgrade,
11: optional i32 avg_score,
12: optional string adjustment
}

struct TableRatingByReport {
1: required list<string> head,
2: required list<RatingByReport> body,
3: optional struct.CadaTranslation translation
}



service china{
void ping(),
TableRatingByReport rating_byreport(1:string ticker) throws (1:struct.CadaInternalError error)
}
Here is my server side:
handler = StockChinaHandler()
processor = china.Processor(handler)
#startup()

transport = TSocket.TServerSocket(port=30303)
tfactory = TTransport.TBufferedTransportFactory()
pfactory = TBinaryProtocol.TBinaryProtocolFactory()

server = TServer.TSimpleServer(processor, transport, tfactory, pfactory)
#server = TProcessPoolServer.TProcessPoolServer(processor, transport,
# tfactory, pfactory)

print "Start server..."
import cProfile
print >>open('/tmp/test.log', 'w'), cProfile.run('server.serve()',
sort='cumulative')
#server.serve()
print "done!"
Client side:
# Make socket
transport = TSocket.TSocket('localhost', 30303)

# Buffering is critical. Raw sockets are very slow
transport = TTransport.TBufferedTransport(transport)

# Wrap in a protocol
protocol = TBinaryProtocol.TBinaryProtocol(transport)

# Create a client to use the protocol encoder
client = china.Client(protocol)

# Connect!
transport.open()

client.ping()
print "ping()"

print msg
msg = client.rating_byreport('2012-01-04')
print msg
transport.close()
cProfile result:
ncalls tottime percall cumtime percall filename:lineno(function)
1 0.000 0.000 230.968 230.968 <string>:1(<module>)
1 0.000 0.000 230.968 230.968 TServer.py:74(serve)
3 0.000 0.000 225.967 75.322 TSocket.py:172(accept)
3 0.000 0.000 225.967 75.322 socket.py:194(accept)
3 225.967 75.322 225.967 75.322 {method 'accept' of '_socket.socket' objects}
5 0.003 0.001 4.993 0.999 china.py:140(process)
1 0.000 0.000 3.200 3.200 china.py:177(process_rating_byreport)
1 0.000 0.000 2.366 2.366 china.py:500(write)
1 0.003 0.003 2.366 2.366 ttypes.py:515(write)
1455 0.261 0.000 2.363 0.002 ttypes.py:364(write)
155556 0.246 0.000 1.995 0.000 TCompactProtocol.py:38(nested)
145880 0.298 0.000 1.640 0.000 TCompactProtocol.py:255(__writeString)
18 1.370 0.076 1.370 0.076 {method 'recv' of '_socket.socket' objects}
5 0.000 0.000 1.292 0.258 TCompactProtocol.py:306(readMessageBegin)
13 0.000 0.000 1.292 0.099 TCompactProtocol.py:286(__readUByte)
26 0.000 0.000 1.291 0.050 TTransport.py:54(readAll)
26 0.000 0.000 1.291 0.050 TTransport.py:154(read)
5 0.000 0.000 1.291 0.258 TSocket.py:101(read)
In my case, TableRatingByReport instance got a body with 1400 rows(list\), and It cost over 3 seconds(function *process_rating_byreport*, which is auto generate by thift) to generate binnary content. I don't know why it is so slow.

Using json to serialize the same data, It's only spend less than 200 ms.

I'm wondering did I use the uncorrect way to manipulate thrift?

Thanks.