Remove delta comparison for fillup (#8)

It is better to let the DB compare the level to the
capacity (using it's own float representation) and to
return based on that. Since we use LEAST the value would be
equal (exactly) to capacity at overflow.

Closes #7

CHANGELOG.md

@@ -1,5 +1,6 @@
 ## [Unreleased]
 
+- [8](https://github.com/cheddar-me/pecorino/pull/8) - Use comparisons in SQL to determine whether the leaky bucket did overflow
 - [6](https://github.com/cheddar-me/pecorino/pull/6) - Changed the way Structs are defined, this does not impact the API.
 
 ## [0.1.0] - 2023-10-30

lib/pecorino/leaky_bucket.rb

@@ -178,15 +178,19 @@ def add_tokens(n_tokens)
               t.level + :fillup - (EXTRACT(EPOCH FROM (EXCLUDED.last_touched_at - t.last_touched_at)) * :leak_rate)
           )
         )
-      RETURNING level
+      RETURNING
+        level,
+        -- Compare level to the capacity inside the DB so that we won't have rounding issues
+        level >= :capa AS did_overflow
     SQL
 
     # Note the use of .uncached here. The AR query cache will actually see our
-    # query as a repeat (since we use "select_value" for the RETURNING bit) and will not call into Postgres
+    # query as a repeat (since we use "select_one" for the RETURNING bit) and will not call into Postgres
     # correctly, thus the clock_timestamp() value would be frozen between calls. We don't want that here.
     # See https://stackoverflow.com/questions/73184531/why-would-postgres-clock-timestamp-freeze-inside-a-rails-unit-test
-    level_after_fillup = conn.uncached { conn.select_value(sql) }
+    upserted = conn.uncached { conn.select_one(sql) }
+    capped_level_after_fillup, did_overflow = upserted.fetch("level"), upserted.fetch("did_overflow")
 
-    State.new(level_after_fillup, (@capacity - level_after_fillup).abs < 0.01)
+    State.new(capped_level_after_fillup, did_overflow)
   end
 end