Saturday, February 28, 2009

JSR-310 javax.time Periods

The proposed JSR-310 date/time API comes with many representations of date and time including Instant, Duration, LocalDate, LocalTime, LocalDateTime, OffsetDate, OffsetTime, OffsetDateTime, ZonedDateTime, MonthDay, YearMonth and others. One of the more interesting classes is Period. It represents a quantity of time, not fixed to any time in space--just a quantity. This entry will discuss periods, how they are parsed, and examples of use.

Parsing: Periods are parsed from string using formats that conform to the ISO-8601 duration format PnYnMnDTnHnMn.nS. Variations of this format are parsed to create Period objects. Parsing is done in PeriodParser, a standalone helper class that is easily accessed through the static method Period.parse(). Here are a few examples:

assert Period.parse("PT0S") == Period.ZERO
assert Period.parse("P1Y") == Period.years(1)
assert Period.parse("P10Y8M22DT3M") == Period.period(10, 8, 22, 3)
assert Period.parse("PT1M") == Period.minutes(1)
assert Period.parse("P-4Y") == Period.years(-4)

As you can see, the parsing scheme is robust and flexible. The main thing to keep in mind is that the Period object is more of a value container as opposed to a process unit. So, 60 minutes won't directly translate to 1 hour, and 24 months does not directly equal 2 years. But this aside, there are many uses for the Period class.

For example, lets say I want to periodically archive old temp files from a system. I could set the period to minutes, hours, days, whatever, then based on an arbitrary start time, begin the archive sweep. In the same moment, I could compute the next archive sweep by adding my pre-defined period object to the current time.

A more elaborate implementation would be to self-adjust the scheduled periods based on some criteria. Sticking with the archive sweep, lets say I set the initial period to 2 hours, or better yet, 120 minutes to give a finer resolution. Then, at the end of the sweep, I tally archived files. The smaller the number, the longer the period, on a sliding scale, up to 240 minutes. A large number would decrease the period to zero minutes, or a continual sweep, so I want to make sure that this is the very worst case scenario.

The Period class makes this easy to implement and maintain. To keep it simple, lets just use a linear equation. The formula for calculating the number of minutes between sweeps is reduced to period in minutes = mx + b, where b = 240 minutes (our maximum amount), x = the number of files, and m = the slope. Lets say that the maximum number anticipated required archives were 2 per minute, so after 240 minutes we would have 480 files that needed to be archived. If that is our worst case then m = -0.5 (delta y divide delta x, or -240 / 480). So the curve looks like this:

Now each time I do an archive sweep, I tally the files and calculate the next sweep interval using:

maxMinutes = 240
slope = -0.5
periodInMinutes = slope * fileCount + maxMinutes

Or, as a single groovy closure:

def periodToNextSweep = { fileCount, slope = -0.5, maxMinutes = 240 ->
Period.minutes( slope * fileCount + maxMinutes )

You might think there is a danger in allowing the returned period to be a negative number of minutes. But in all practicality this is acceptable because the objective is to determine the next instant when a sweep should occur. If this time is in the past, simply do it now. Of coarse you would design the slope to target the worst case, so a negative time should seldom if ever occur. And the good part is that the parameters are easy to modify to fit changing environments.

Testing with a Fixed Time Source: My previous entries have demonstrated using TimeSource and Clock tied to the System clock. But, for these tests, I think a fixed time source would be more appropriate. The syntax is like this:

millis = 1234920035991L // 2009-02-17T17:20:35.991-08:00 Tuesday...
timeSource = TimeSource.fixed( Instant.millisInstant( millis ) )
clock = Clock.clockDefaultZone( timeSource )

So now when I create a date, time or date/time object from the clock, the time is always the same. Not very meaningful for real life, but great for testing.

If I create a class that uses clock, I can inject the TimeSource based on the system clock. And for testing, I can inject a fixed TimeSource, run tests, and not have to worry about the specific time, but simply base my tests on a static source. Here is the class:

class SweepController {
def clock = Clock.systemDefaultZone()
def slope = -0.5
def maxMinutes = 240

def periodToNextSweep = { fileCount ->
int x = (int)(slope * fileCount + maxMinutes)
x < 0 ? Period.ZERO : Period.minutes( x )

def nextSweepTime = { fileCount ->
def period = periodToNextSweep( fileCount )

clock.offsetDateTime() + period

And here is the test script:

millis = 1235721600000L // 2009-02-27T00:00-08:00
fixed = TimeSource.fixed( Instant.millisInstant( millis ) )
clock = Clock.clockDefaultZone( fixed )

sweep = new SweepController( clock:clock )
println "now -> ${clock.offsetDateTime()}"
source = [
[ 480, '2009-02-27T00:00-08:00' ],
[ 0, '2009-02-27T04:00-08:00' ],
[ 240, '2009-02-27T02:00-08:00' ],
[ 120, '2009-02-27T03:00-08:00' ],
source.each { count, value ->
println "count: ${count} -> ${sweep.periodToNextSweep( count )}, ${sweep.nextSweepTime( count )}"
assert value == sweep.nextSweepTime( count ).toString()

When I run the script, here is what I get:

now -> 2009-02-27T00:00-08:00
count: 480 -> PT0S, 2009-02-27T00:00-08:00
count: 0 -> PT240M, 2009-02-27T04:00-08:00
count: 240 -> PT120M, 2009-02-27T02:00-08:00
count: 120 -> PT180M, 2009-02-27T03:00-08:00

The main advantage is that I can run this independent of the current date, but still use the clock object without changing anything inside the class.

Conclusion: This quick look at Period and PeriodParser to see how it fits into the JSR-310 from the groovy coder's perspective. Next time well look closer at Date, Time and DateTime math capabilities and how they work with groovy plus/minus operator overloading.

No comments: