You Might Not Want Async

First version given at PyCon TW 2016.
Revised for PyCon APAC 2016 (Seoul).
Revised for PyCon JP 2016.

Asynchrony in Python had gathered much momentum recently, with interests from core developers, as evidenced by the introduction of `asyncio` in Python 3.4, and a great boom of related third-party projects following it. By utilising more functionalities from the underlying operating system, it is a great solution to many existing problems in Python applications, gaining practical concurrency without working around the well-known GIL (global interpreter lock) problem.

With all its advantages, asynchrony is, however, still a relatively new concept in Python, and as a result could be somewhat mistaken, even misunderstood by some people. One of these misconceptions, probably the most serious, is to mistake concurrency through asynchrony for parallelism. Although `asyncio` (and other similar solutions) lets multiple parts of your program executes without interfering each other, it does *not* allow them to run together—this is still impossible, at least in CPython, due to the continued existence of the GIL. This makes asynchrony suitable for only a certain, instead of all, kinds of problems. Evaluation is therefore required before a programmer can decide whether the asynchrony model is suitable for a particular application.

Furthermore, partly due to its relatively short existence, paradigms in asynchrony programming do not necessarily fit well with other parts of Python, including libraries, either built-in or third-party ones. Since only blocking libraries were available in most of Python’s history, many assumptions they made may not work well with async programs out-of-the-box. Adopting asynchrony, at least at the present time, will therefore introduce more technical debt to your program. These are all important aspects that require much consideration before you dive head-first into asynchrony.

Transcript

1. You Might Not Want Async
   

   View Slide

2. Quick Questions
   • Concurrency with Python
   • Threads
   • Multi-processing
   • Single-thread asynchrony
   • asyncio
   

   View Slide

3. ࣗݾ঺հ (PyCon JP Ver.)
   • ৉ ࢠሯ
   • ͫΐ͏ ͪʔͽΜ
   • @uranusjr
   

   View Slide

4. Me
   • Call me TP
   • Follow @uranusjr
   • https://uranusjr.com
   

   View Slide

5. View Slide

6. View Slide

7. http://macdown.uranusjr.com
   

   View Slide

8. www. .com
   

   View Slide

9. 10–11 June 2017 (Ծ)
   7
   

   View Slide

10. View Slide

11. View Slide

12. meh
    

    View Slide

13. 2014
    Python 3.4 OSDC.tw PyCon APAC
    March April May
    

    View Slide

14. View Slide

15. Got me
    thinking
    

    View Slide

16. View Slide

17. Synchronous
    

    View Slide

18. (Single-Threaded) Async
    

    View Slide

19. Sync vs. Async
    

    View Slide

20. Absolute
    magic
    

    View Slide

21. Before After
    

    View Slide

22. Module 5 - Doctor Faustus by Christopher Marlowe
    

    View Slide

23. Not For You
    • Infects the whole program
    • Async is not parallelism
    • Third party support
    

    View Slide

24. import sqlite3
    def read_data(dbname):
    con = sqlite3.connect(dbname)
    cur = con.cursor()
    cur.execute('SELECT * FROM data OFFSET 0 LIMIT 1')
    data = cur.fetchone()
    cur.close()
    con.close()
    return data
    

    View Slide

25. import aioodbc
    async def read_data(dbname):
    con = await aioodbc.connect(
    dsn='Driver=SQLite;Database={}'.format(dbname),
    )
    cur = await con.cursor()
    await cur.execute('SELECT * FROM data OFFSET 0 LIMIT 1')
    data = await cur.fetchone()
    await cur.close()
    await con.close()
    return data
    

    View Slide

26. from .db import read_data
    def main():
    data = read_data('data.sqlite3')
    print(data)
    main()
    

    View Slide

27. import asyncio
    from .db import read_data
    async def main():
    data = await read_data('db.sqlite3')
    print(data)
    loop = asyncio.get_event_loop()
    loop.run_until_complete(main())
    loop.close()
    

    View Slide

28. U+1F937 SHRUG (Unicode 9.0)
    

    View Slide

29. RELAX PEOPLE
    I’M JUST GETTING STARTED
    

    View Slide

30. import asyncio
    from .db import read_data
    async def main():
    data = read_data('db.sqlite3')
    print(data)
    loop = asyncio.get_event_loop()
    loop.run_until_complete(main())
    loop.close()
    

    View Slide

31. (demo) $ python demo.py
    
    

    View Slide

32. import asyncio
    from .db import read_data
    async def main():
    data = read_data('db.sqlite3')
    print(data)
    loop = asyncio.get_event_loop()
    loop.run_until_complete(main())
    loop.close()
    

    View Slide

33. import asyncio
    from .db import read_data
    async def main():
    data = await read_data('db.sqlite3')
    print(data)
    loop = asyncio.get_event_loop()
    loop.run_until_complete(main())
    loop.close()
    

    View Slide

34. View Slide

35. I KNOW
    WRITE UNIT TESTS
    

    View Slide

36. GOOD LUCK
    WITH THAT
    

    View Slide

37. View Slide

38. View Slide

39. import asyncio
    import time
    async def do_something():
    print('Before', time.monotonic())
    await asyncio.sleep(2)
    print('After ', time.monotonic())
     await do_something()
    Before 199208.190632853
    After 199210.192092425
    

    View Slide

40. import unittest
    class MyTestCase(unittest.TestCase):
    async def test_do_something(self):
    await do_something()
    if __name__ == '__main__':
    unittest.main()
    

    View Slide

41. (demo) $ python tests.py
    .
    ----------------------------------------------------
    Ran 1 test in 0.002s
    OK
    

    View Slide

42. (demo) $ python tests.py
    .
    ----------------------------------------------------
    Ran 1 test in 0.002s
    OK
    

    View Slide

43. What Happened?
    • unittest does not know about asyncio
    • Coroutine methods are executed “normally”
    • Called, but not executed (awaited)
    

    View Slide

44. WHAT IF I TOLD YOU
    THIS IS GONNA BE TOUGH
    

    View Slide

45. import asyncio
    import functools
    def asynchronous(func):
    @functools.wraps(func)
    def asynchronous_inner(*args, **kwargs):
    loop = asyncio.new_event_loop()
    asyncio.set_event_loop(loop)
    try:
    loop.run_until_complete(func(*args, **kwargs))
    finally:
    loop.close()
    return asynchronous_inner
    

    View Slide

46. import unittest
    class MyTestCase(unittest.TestCase):
    @asynchronous
    async def test_do_something(self):
    await do_something()
    if __name__ == '__main__':
    unittest.main()
    

    View Slide

47. (demo) $ python tests.py
    Before 51728.420457105
    After 51730.424515145
    .
    ----------------------------------------------------
    Ran 1 test in 2.006s
    OK
    

    View Slide

48. Tips
    • Beware of warning output
    • Especially if you redirect
    • Add coverage report to testing code
    • Consider pip install asynctest
    

    View Slide

49. Or use pytest instead
    

    View Slide

50. import time
    async def test_do_something():
    before = time.monotonic()
    await do_something()
    delta_t = time.monotonic() - before
    assert -0.01 < delta_t < 0.01
    

    View Slide

51. (demo) $ py.test tests.py
    ====================== test session starts ======================
    platform darwin -- Python 3.5.1, pytest-2.9.1, py-1.4.31,
    pluggy-0.3.1
    collected 0 items / 1 errors
    ============================ ERRORS =============================
    ___________________ ERROR collecting tests.py____________________
    > for i, x in enumerate(self.obj()):
    E TypeError: 'coroutine' object is not iterable
    python3.5/site-packages/_pytest/python.py:765: TypeError
    ==================== 1 error in 0.15 seconds ====================
    

    View Slide

52. View Slide

53. import time
    import pytest
    @pytest.mark.asyncio
    async def test_do_something(capsys):
    before = time.monotonic()
    await do_something()
    delta_t = time.monotonic() - before
    assert -0.01 < delta_t < 0.01
    

    View Slide

54. View Slide

55. View Slide

56. View Slide

57. asyncio
    • Boilerplate
    • Error-prone
    • Immature API (I think)
    

    View Slide

58. import requests
    def collect_contents(urls):
    contents = []
    for url in urls:
    resp = requests.get(url)
    if resp.status_code != 200:
    continue
    content = resp.text
    contents.append(content)
    return contents
    

    View Slide

59. import aiohttp
    async def collect_contents(urls):
    contents = []
    with aiohttp.ClientSession() as session:
    for url in urls:
    async with session.get(url) as rest:
    if resp.status != 200:
    continue
    content = await resp.text()
    contents.append(content)
    return contents
    

    View Slide

60. 㖶װ׌ַ׵ׁ
    

    View Slide

61. import aiohttp
    async def collect_contents(urls):
    contents = []
    with aiohttp.ClientSession() as session:
    for url in urls:
    async with session.get(url) as rest:
    if resp.status != 200:
    continue
    content = await resp.text()
    contents.append(content)
    return contents
    

    View Slide

62. View Slide

63. import asyncio
    import aiohttp
    async def collect_contents(urls):
    coroutines = []
    with aiohttp.ClientSession() as session:
    for url in urls:
    async with session.get(url) as rest:
    if resp.status != 200:
    continue
    coroutines.append(resp.text())
    contents = await asyncio.gather(*coroutines)
    return contents
    

    View Slide

64. View Slide

65. View Slide

66. View Slide

67. LET’S TRY SOMETHING
    “DIFFERENT”
    

    View Slide

68. Alternatives
    • concurrent & multiprocessing
    • Greenlets
    • Similar idea, but less infectious
    • C extension
    • threading
    • Standard I/O
    

    View Slide

69. http://greenlet.readthedocs.io
    

    View Slide

70. GEVENT
    MEH
    

    View Slide

71. ROUTINES
    ROUTINES EVERYWHERE
    

    View Slide

72. package main
    import ("fmt"; "time")
    func doSomething() {
    time.Sleep(2 * time.Second)
    fmt.Println(time.Now(), "Slept")
    }
    func main() {
    doSomething()
    fmt.Println(time.Now(), "OK")
    }
    00:00:00 Slept
    00:00:00 OK
    

    View Slide

73. package main
    import ("fmt"; "time")
    func doSomething() {
    time.Sleep(2 * time.Second)
    fmt.Println(time.Now(), "Slept")
    }
    func main() {
    go doSomething()
    fmt.Println(time.Now(), "OK")
    }
    00:00:00 OK
    

    View Slide

74. package main
    import ("fmt"; "time")
    var sem = make(chan bool)
    func doSomething() {
    time.Sleep(2 * time.Second)
    fmt.Println(time.Now(), "Slept")
    sem  true
    }
    func main() {
    go doSomething()
    fmt.Println(time.Now(), "OK")
    sem
    }
    00:00:00 OK
    00:00:02 Slept
    

    View Slide

75. The Go Model
    • Boilerplate
    • Error-prone
    • Immature API (I think)
    

    View Slide

76. """Do something. Synchronous version."""
    import datetime
    import time
    def do_something():
    time.sleep(2)
    print(datetime.datetime.now(), 'Slept')
    do_something()
    print(datetime.datetime.now(), 'Done')
    

    View Slide

77. """Do something. Asynchronous version."""
    import asyncio
    import datetime
    async def do_something():
    await asyncio.sleep(2)
    print(datetime.datetime.now(), 'Slept')
    loop = asyncio.get_event_loop()
    task = loop.create_task(do_something())
    print(datetime.datetime.now(), 'Done')
    loop.run_until_complete(asyncio.wait([task]))
    loop.close()
    

    View Slide

78. View Slide

79. """Do something. Synchronous version."""
    import datetime
    import time
    def do_something():
    time.sleep(2)
    print(datetime.datetime.now(), 'Slept')
    do_something()
    print(datetime.datetime.now(), 'Done')
    

    View Slide

80. """What if I can just write this?"""
    import asyncio
    import datetime
    import time
    async def do_something():
    await time.sleep(2)
    print(datetime.datetime.now(), 'Slept')
    await do_something()
    print(datetime.datetime.now(), 'Done')
    asyncio.run_event_loop()
    

    View Slide

81. I know, it’s not really
    possible.
    

    View Slide

82. At least we can dream.
    Or wait until Python 6.0?
    

    View Slide

83. Recap
    • Rant
    • Moar rant
    • Susceptible advice
    • Unrealistic dream
    

    View Slide

84. But Seriously
    • Asynchrony is not the silver bullet
    • It makes you jump through loops
    • There are alternatives
    • Fingers crossed
    

    View Slide

85. View Slide