求一个关于日期时间间隔计算相关的算法,具体问题如下:
当前有一个比较大(上万个)日期时间的list(时间递增但无规律),假设 dt=[
'2015-01-02 10:21:02',
'2015-01-02 10:21:02'
'2015-01-02 10:21:03'
'2015-01-02 10:21:04'
'2015-01-02 10:21:06'
'2015-01-02 10:21:10'
'2015-01-02 10:21:11'
'2015-01-02 10:21:13'
'2015-01-02 10:22:00'
'2015-01-02 10:22:12'
... ...
]
给定一个时间间隔timeframe(假设4s),现在问题是:求在给定连续时间间隔内,list中最多项数目。
最简单的方法是依次对于list中每个项x,向后查找直至大于timeframe+x的项并统计个数。但由于这个list数量庞大,同时还要涉及时间运算,算法难以达到要求,特来请教!
大家讲道理2017-04-18 09:23:33
先給我的答案:
def findmax(timestrs, timeframe):
fmt = "%Y-%m-%d %H:%M:%S"
times = [datetime.strptime(timestr, fmt) for timestr in timestrs]
frame = timedelta(seconds=timeframe)
window = dict(bidx=0, eidx=0)
maxwindow = dict(bidx=0, eidx=0)
# search
for idx, time in enumerate(times):
if time-times[window['bidx']] > frame:
count = window['eidx'] - window['bidx'] + 1
if count > maxwindow['eidx'] - maxwindow['bidx']:
maxwindow = dict(window)
while time-times[window['bidx']] > frame:
window['bidx'] += 1
window['eidx'] = idx
else:
count = window['eidx'] - window['bidx'] + 1
if count > maxwindow['eidx'] - maxwindow['bidx']:
maxwindow = dict(window)
# output results
maxcount = maxwindow['eidx'] - maxwindow['bidx'] + 1
print('max count:{} of winodw from {}:{} to {}:{}'.format(
maxcount, maxwindow['bidx'], timestrs[maxwindow['bidx']],
maxwindow['eidx'], timestrs[maxwindow['eidx']]))
上述的 code 要記得 import datetime
和 timedelta
.
這是一個 O(n) 的做法, n 為 timestrs
的數量.
code 有點醜, 晚一點再修, 不過為了比較速度跟正確性, 我做了一些實驗, 以下是我用來做實驗完整的 code:
from datetime import datetime, timedelta
import collections
import random
import sys
def simpletimer(func):
""" a simple decorator to time a function"""
import time
def modified_func(*args, **kwargs):
btime = time.time()
result = func(*args, **kwargs)
print('[time]', time.time()-btime, 'sec.')
return result
return modified_func
def gentimestrs(count, delta):
""" generate time strings for input data"""
timestrs = []
time = datetime(2015, 1, 2)
for i in range(count):
time = time + timedelta(seconds=random.randrange(delta))
timestrs.append(str(time))
return timestrs
@simpletimer
def findmax(timestrs, timeframe):
fmt = "%Y-%m-%d %H:%M:%S"
times = [datetime.strptime(timestr, fmt) for timestr in timestrs]
frame = timedelta(seconds=timeframe)
window = dict(bidx=0, eidx=0)
maxwindow = dict(bidx=0, eidx=0)
for idx, time in enumerate(times):
if time-times[window['bidx']] > frame:
count = window['eidx'] - window['bidx'] + 1
if count > maxwindow['eidx'] - maxwindow['bidx']:
maxwindow = dict(window)
while time-times[window['bidx']] > frame:
window['bidx'] += 1
window['eidx'] = idx
else:
count = window['eidx'] - window['bidx'] + 1
if count > maxwindow['eidx'] - maxwindow['bidx']:
maxwindow = dict(window)
maxcount = maxwindow['eidx'] - maxwindow['bidx'] + 1
print('max count:{} of winodw from {}:{} to {}:{}'.format(
maxcount, maxwindow['bidx'], timestrs[maxwindow['bidx']],
maxwindow['eidx'], timestrs[maxwindow['eidx']]))
@simpletimer
def findmax2(timestrs, timeframe):
fmt = "%Y-%m-%d %H:%M:%S"
ys = ((datetime.strptime(x, fmt) + timedelta(seconds=-i)).strftime(fmt) for x in timestrs for i in range(timeframe))
print(collections.Counter(ys).most_common(1))
if __name__ == '__main__':
count = int(sys.argv[1])
delta = int(sys.argv[2])
frame = int(sys.argv[3])
timestrs = gentimestrs(count, delta)
with open('timestrs', 'w') as writer:
for timestr in timestrs:
print(timestr, file=writer)
funclst = [findmax, findmax2]
for func in funclst:
func(timestrs, frame)
包含了
一個很陽春用來計時的 decorator: simpletimer
他會裝飾 func
, 印出執行所花的時間
一個產生隨機測資的 function: gentimestrs
他會產生 count
筆 time strings 測資, 且各資料的時間間隔是小於等於 delta
的隨機整數
接下來的findmax
和findmax2
分別是我的做法跟@citaret 的做法(有興趣的大家可以自行仿照界面加入自己的function, 題主也可以把自己原先的做法拿來比比看), 並且用findmax
和 findmax2
分別是我的做法跟 @citaret 的做法 (有興趣的大家可以自行仿照介面加入自己的 function, 題主也可以把自己原先的做法拿來比比看), 並且用 simpletimer
計時.
測試的時候先產生測資, 再分別執行 funclst
中要比較的 funcion
執行的範例如下:
timeframe
_
$ python3 test.py 10000 30 6
^^^^^ ^^
測資數 隨機最大時間間隔
這邊列出一些結果:
用 timeframe==5
測試測資數量 1000, 10000, 100000
$ python3 td.py 1000 30 5
max count:4 of winodw from 798:2015-01-02 03:25:56 to 801:2015-01-02 03:26:01
[time] 0.02829718589782715 sec.
[('2015-01-02 03:25:56', 3)]
[time] 0.15825390815734863 sec.
$ python3 td.py 10000 30 5
max count:5 of winodw from 1624:2015-01-02 06:37:32 to 1628:2015-01-02 06:37:37
[time] 0.19979143142700195 sec.
[('2015-01-02 12:17:22', 4)]
[time] 1.6265528202056885 sec.
$ python3 td.py 100000 30 5
max count:6 of winodw from 91688:2015-01-17 09:41:33 to 91693:2015-01-17 09:41:38
[time] 1.8956780433654785 sec.
[('2015-01-15 01:36:41', 5)]
[time] 15.950862407684326 sec.
用 timeframe==10
測試測資數量 1000, 10000, 100000
$ python3 td.py 1000 30 10
max count:5 of winodw from 910:2015-01-02 03:36:08 to 914:2015-01-02 03:36:17
[time] 0.02788376808166504 sec.
[('2015-01-02 03:24:06', 5)]
[time] 0.3295907974243164 sec.
$ python3 td.py 10000 30 10
max count:6 of winodw from 5304:2015-01-02 21:45:41 to 5309:2015-01-02 21:45:47
[time] 0.20185112953186035 sec.
[('2015-01-02 21:45:38', 6)]
[time] 3.2009713649749756 sec.
$ python3 td.py 100000 30 10
max count:7 of winodw from 16224:2015-01-04 17:16:17 to 16230:2015-01-04 17:16:23
[time] 1.8946728706359863 sec.
[('2015-01-04 17:16:17', 7)]
[time] 33.85069417953491 sec.
若維持 timeframe 大小不變, 兩者的時間對測資數量都呈現線性成長
可是當 timeframe 變成兩倍時, 第二種做法的時間也會多一倍
這表示 @citaret 大的方法複雜度應該是 O(n * timeframe)
還有一點就是有時候這兩個做法的結果不太相同, 會差一, 題主若有方法可以驗證一下正確性
我單單看我找出來的答案似乎沒錯, 但若有疏忽還請大家指正我, 謝謝!
也歡迎更多人討論這題.
我回答過的問題: Python-QA
高洛峰2017-04-18 09:23:33
沒有數據,不好 profile,先給一版,看看別人是不是有更好的處理。
xs = [
'2015-01-02 10:21:02',
'2015-01-02 10:21:02',
'2015-01-02 10:21:03',
'2015-01-02 10:21:04',
'2015-01-02 10:21:06',
'2015-01-02 10:21:10',
'2015-01-02 10:21:11',
'2015-01-02 10:21:13',
'2015-01-02 10:22:00',
'2015-01-02 10:22:12' ]
from datetime import datetime, timedelta
import collections
delta = 5
fmt = "%Y-%m-%d %H:%M:%S"
ys = ((datetime.strptime(x, fmt) + timedelta(seconds=-i)).strftime(fmt) for x in xs for i in range(delta))
print collections.Counter(ys).most_common(1)
# output: [('2015-01-02 10:21:02', 5)]
伊谢尔伦2017-04-18 09:23:33
@dokelung 使用citaret演算法測試了下,對於共27505個元素的list(從有63萬行單一日誌檔案中提取,多個來源ip匯總)耗時3.8842415850296845s,我原來時耗時16.544873371035163s,不少,雖然數量級還在秒,但這只是一天的一個文件。同時幫我發現效率問題更嚴重的其實是在函數內部對滿足要求日誌行的提取上,感謝大家的回答與參與,謝謝!