Thread-safe java.text.SimpleDateFormat format and parse

[Claes Redestad]

+1

The suggested approaches would have a negative performance
impact on code which have addressed the thread-safety issues
(by synchronizing, using ThreadLocal instances or similar), so I see
no way this could be fixed in a way that would make everyone
happy.

/Claes

On 2015-07-01 17:39, Ben Evans wrote:
> Given the other problems with the legacy date and time classes, why
> spend engineering time tidying this up?
>
> Everyone should be migrating to the new date & time support in
> java.time, so this would just be a distraction.
>
> Thanks,
>
> Ben
>
> On Sat, Jun 27, 2015 at 8:36 AM, Paul Draper <paulddraper at gmail.com> wrote:
>> While it's often understood that SimpleDateFormat isn't thread safe with
>> its setters, etc. it is frequently incorrectly assumed (despite the docs)
>> that since format() and parse() do not mutate the object in a visible way,
>> they can be called from multiple threads.
>>
>> The rationale is akin to calling ArrayList#get or HashMap#get from multiple
>> threads. The entire class is not thread-safe, but you can call that
>> non-mutating accessor from multiple threads without issue.
>> The trouble is that SimpleDateFormat has a private Calendar instance
>> variable, which is mutated during the format() and parse() methods.
>>
>> This is a very common mistake. There is a project whose entire purpose is a
>> thread-safe formatter: https://code.google.com/p/safe-simple-date-format/
>> And Apache Commons and Joda Time provide similar classes.
>>
>> Currently, users of SimpleDateFormat have to synchronize format() and
>> parse(), or use a separate SimpleDateFormat for every thread.
>> Or, too commonly, do neither and have a relatively unobvious race condition.
>>
>> Making format() and parse() calls thread-safe would require either using a
>> local Calendar variable -- one instance per call -- or using a thread-local
>> Calendar -- one instance per thread. The former option seems the best.
>>
>> The change would be fully backwards compatible. I have profiled a change
>> with a local Calendar variable, and measured no difference in the
>> performance (format and parse are by their nature rather involved methods
>> to begin with).
>> This change would improve the intuitive behavior of SimpleDateFormat and
>> eliminate one of the most common mistakes of JDK users.