Timestamp collisions demo for this blog post: http://randnotizen.de/post/18898864701/will-openstack-swift-lose-data

collisions.py

import time

def test_time():
    collisions = 0  # number of collisions
    dist = {}       # distinct values
    microsecond = 0.000001
    count = 0
    while count < 1000000:
        count += 1
        t1 = time.time()
        time.sleep(microsecond)
        t2 = time.time()
        s1 = "%016.05f" % t1
        s2 = "%016.05f" % t2
        if s1 == s2:
            collisions += 1
        else:
            diff = t2 - t1
            dist[diff] = dist.get(diff, 0) + 1
    percent = collisions / (count / 100.0)
    print "Collisions:\t%s (%s%%)" % (collisions, percent)
    print "Min value:\t%s" % min(dist.iterkeys())
    print "Max value:\t%s" % max(dist.iterkeys())
    print "Dist values:\t%s" % len(dist)
    

if __name__ == '__main__':
    test_time()

Cool on the disqus thing, we should probably move over to that. :)

What exactly happens depends somehwat upon the number of replicas Swift is configured for. If we assume the standard 3 replicas, you could theoretically end up with 3 different "answers" to the question: What is the data in this object? But, since all three were at the exact same collision time, there's really no way to know which is most correct anyway.

Many systems use a content hash to resolve these conflicts at sync, but early in Swift's development it was decided that highly concurrent single object writes weren't a use case for Swift needing such extra resolution steps.

For Swift, it would keep those multiple versions until the hardware on one failed, in which case it would get one of the other versions, etc. until there was just one version again.