forked from alibaba/jstorm
-
Notifications
You must be signed in to change notification settings - Fork 0
/
changehistory
112 lines (100 loc) · 5.31 KB
/
changehistory
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
# Release 0.9.1
1. Application classloader. when Application jar is conflict with jstorm jar,
please enable application classloader.
2. Group Quato, Different group with different resource quato.
Bug fix
1. Fix Rotation Map competition issue.
2. Set default acker number as 0
3. Set default spout/bolt number as 1
4. Add log directory in log4j configuration file
5. Add transaction example
6. Fix UI showing wrong worker numbe in topology page
7. Fix UI showing wrong latency in topology page
8. Replace hardcode Integer convert with JStormUtils.parseInt
9. Support string parse in Utils.getInt
10. Remove useless dependency in pom.xml
11. Support supervisor using IP or special hostname
12. Add more details when no resource has been assigned to one new topology
13. Replace normal thread with Smart thread
14. Add gc details
15. Code format
16. Unify stormId and topologyId as topologyId
17. Every nimbus will regist ip to ZK
# Release 0.9.0
In this version, it will follow storm 0.9.0 interface, so the application running
on storm 0.9.0 can run in jstorm 0.9.0 without any change.
Stability
1. provide nimbus HA. when the master nimbus shuts down, it will select another
online nimbus to be the master. There is only one master nimbus online
any time and the slave nimbuses just synchronouse the master's data.
2. RPC through netty is stable, the sending speed is match with receiving speed.
Powerful scheduler
1. Assigning resource on four dimensions:cpu, mem, disk, net
2. Application can use old assignment.
3. Application can use user-define resource.
4. Task can application extra cpu slot or memory slot.
4. Application can force tasks run on different supervisor or the same supervisor
# Release 0.7.1
In this version, it will follow storm 0.7.1 interface, so the topology running
in storm 0.7.1 can run in jstorm without any change.
Stability
1. add setting "zmq.max.queue.msg" for zeromq
2. communication between worker and tasks without zeromq
3. Add catch exception operation
3.1. in supervisor SyncProcess/SyncSupervisor
3.2. add catch exception and report_error in spout's open and bolt's prepare
3.3. in all IO operation
3.4. in all serialize/deserialize
3.5. in all ZK operation
3.6 in topology upload/download function
3.7 during initialization zeromq
4. do assignmen/reassignment operation in one thread to avoid competition
5. redesign nimbus 's topology assign algorithm, make the logic simple much.
6. redesign supervisor's sync assignment algorithm, make the logic simple much
7. reduce zookeeper load
7.1 redesign nimbus monitor logic, it will just scan tasks' hearbeat, frequency is 10s
7.2 nimbus cancel watch on supervisor
7.3 supervisor heartbeat frequence change to 10s
7.4 supervisor syncSupervisor/syncProcess frequence change to 10s
7.5 supervisor scan /$(ZKROOT)/assignment only once in one monitor loop
7.6 task hearbeat change to 10s
8 create task pid file before connection zk, this is very import when zk is unstable.
Performance tuning
1. reduce once memory copy when deserialize tuple, improve performance huge.
2. split executor thread as two thread, one handing receive tuples, one sending tuples, improve performance much
3. redeisign sample code, it will sampling every 5 seconds, not every 20 tuple once, improve performance much
3. simplify the ack's logic, make acker more effeciency
4. Communication between worker and tasks won't use zeromq, just memory share in process
5. in worker's Drainer/virtualportdispatch thread, spout/bolt recv/send thread,
the thread will sleep 1 ms when there is not tuple in one loop
6. communication between worker and tasks without zeromq
7. sampling frequence change to 5s, not every 20 tuple once.
Enhancement:
1. add IFailValueSpout interface
2. Redesign sampling code, collection statics model become more common.
Add sending/recving tps statics, statics is more precise.
3. Atomatically do deactivate action when kill/rebalance topology,
and the wait time is 2 * MSG_TIMEOUT
4. fix nongrouping bug, random.nextInt will generate value less than 0.
5. Sleep one setting time(default is 1 minute) after finish spout open,
which is used to wait other task finish initialization.
6. Add check component name when submit topology, forbidding the component
which name start with "__"
7. change the zk's node /$(ZKROOT)/storm to /$(ZKROOT)/topology
8. abstract topology check logic from generating real topology function
9. when supervisor is down and topology do rebalance, the alive task under down
supervisor is unavailable.
10. add close connection operation after finish download topology binary
11. automatically create all local dirtorie, such as
/$(LOCALDIR)/supervisor/localstate
12. when killing worker, add "kill and sleep " operation before "kill -9" operation
13. when generate real topology binary,
13.1. configuration priority different.
component configuration > topology configuration > system configuration
13.2. skip the output stream which target component doesn't exist.
13.3. skip the component whose parallism is 0.
13.4. component's parallism is less than 0, throw exception.
14. skip ack/fail when inputstream setting is empty
15. add topology name to the log
16. fix ui select option error, default is 10 minutes
17. supervisor can display all worker's status