Introduction to PipelineDuration

1 files changed, 98 insertions, 0 deletions

diff --git a/lib/gitlab/ci/pipeline_duration.rb b/lib/gitlab/ci/pipeline_duration.rb
index 55d870e2e9d..f97727e8548 100644
--- a/lib/gitlab/ci/pipeline_duration.rb
+++ b/lib/gitlab/ci/pipeline_duration.rb
@@ -1,5 +1,103 @@
 module Gitlab
   module Ci
+    # The problem this class is trying to solve is finding the total running
+    # time amongst all the jobs, excluding retries and pending (queue) time.
+    # We could reduce this problem down to finding the union of periods.
+    #
+    # So each job would be represented as a `Period`, which consists of
+    # `Period#first` and `Period#last`. A simple example here would be:
+    #
+    # * A (1, 3)
+    # * B (2, 4)
+    # * C (6, 7)
+    #
+    # Here A begins from 1, and ends to 3. B begins from 2, and ends to 4.
+    # C begins from 6, and ends to 7. Visually it could be viewed as:
+    #
+    #     0  1  2  3  4  5  6  7
+    #        AAAAAAA
+    #           BBBBBBB
+    #                       CCCC
+    #
+    # The union of A, B, and C would be (1, 4) and (6, 7), therefore the
+    # total running time should be:
+    #
+    #     (4 - 1) + (7 - 6) => 4
+    #
+    # And the pending (queue) time would be (4, 6) like this: (marked as X)
+    #
+    #     0  1  2  3  4  5  6  7
+    #        AAAAAAA
+    #           BBBBBBB
+    #                       CCCC
+    #                  XXXXX
+    #
+    # Which could be calculated by having (1, 7) as total time, minus
+    # the running time we have above, 4. The full calculation would be:
+    #
+    #     total = (7 - 1)
+    #     duration = (4 - 1) + (7 - 6)
+    #     pending = total - duration # 6 - 4 => 2
+    #
+    # Which the answer to pending would be 2 in this example.
+    #
+    # The algorithm used here for union would be described as follow.
+    # First we make sure that all periods are sorted by `Period#first`.
+    # Then we try to merge periods by iterating through the first period
+    # to the last period. The goal would be merging all overlapped periods
+    # so that in the end all the periods are discrete. When all periods
+    # are discrete, we're free to just sum all the periods to get real
+    # running time.
+    #
+    # Here we begin from A, and compare it to B. We could find that
+    # before A ends, B already started. That is `B.first <= A.last`
+    # that is `2 <= 3` which means A and B are overlapping!
+    #
+    # When we found that two periods are overlapping, we would need to merge
+    # them into a new period and disregard the old periods. To make a new
+    # period, we take `A.first` as the new first because remember? we sorted
+    # them, so `A.first` must be smaller or equal to `B.first`. And we take
+    # `[A.last, B.last].max` as the new last because we want whoever ended
+    # later. This could be broken into two cases:
+    #
+    #     0  1  2  3  4
+    #        AAAAAAA
+    #           BBBBBBB
+    #
+    # Or:
+    #
+    #     0  1  2  3  4
+    #        AAAAAAAAAA
+    #           BBBB
+    #
+    # So that we need to take whoever ends later. Back to our example,
+    # after merging and discard A and B it could be visually viewed as:
+    #
+    #     0  1  2  3  4  5  6  7
+    #        DDDDDDDDDD
+    #                       CCCC
+    #
+    # Now we could go on and compare the newly created D and the old C.
+    # We could figure out that D and C are not overlapping by checking
+    # `C.first <= D.last` is `false`. Therefore we need to keep both C
+    # and D. The example would end here because there are no more jobs.
+    #
+    # After having the union of all periods, the rest is simple and
+    # described in the beginning. To summarise:
+    #
+    #     duration = (4 - 1) + (7 - 6)
+    #     total = (7 - 1)
+    #     pending = total - duration # 6 - 4 => 2
+    #
+    # Note that the pending time is actually not the final pending time
+    # for pipelines, because we still need to accumulate the pending time
+    # before the first job (A in this example) even started! That is:
+    #
+    #     total_pending = pipeline.started_at - pipeline.created_at + pending
+    #
+    # Would be the final answer. We deal with that in pipeline itself
+    # but not here because here we try not to be depending on pipeline
+    # and it's trivial enough to get that information.
     class PipelineDuration
       PeriodStruct = Struct.new(:first, :last)
       class Period < PeriodStruct