Use better histogram buckets (#8263)

Use a method of generating buckets based on specifying how many
buckets to use per order of magnitude.

For our bytes field, we use:
(32, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072, 262144, 524288, 1048576)

And for querys per connection, we use:
(1.0, 5.0, 10.0, 50.0, 100.0, 500.0, 1000.0, 5000.0, 10000.0)

Fixes #8245.

edb/common/prometheus.py

@@ -70,13 +70,36 @@ def calc_buckets(
     """Calculate histogram buckets on a logarithmic scale."""
     # See https://amplitude.com/blog/2014/08/06/optimal-streaming-histograms
     # for more details.
+    # (Says a long standing comment, but this isn't what that post recommends!)
     result: list[float] = []
     while start <= upper_bound:
         result.append(start)
         start *= increment_ratio
     return tuple(result)
 
 
+def per_order_buckets(
+    start: float, end: float,
+    *,
+    base: float=10.0,
+    entries_per_order=4,
+) -> tuple[float, ...]:
+    # See https://amplitude.com/blog/2014/08/06/optimal-streaming-histograms
+    # for more details.
+    # (Actually, for this one.)
+    result: list[float] = [start]
+
+    next = start * base
+    while next <= end:
+        for i in range(1, entries_per_order):
+            val = next / entries_per_order * i
+            if val > result[-1]:
+                result.append(val)
+        result.append(next)
+        next *= base
+    return tuple(result)
+
+
 class Unit(enum.Enum):
 
     # https://prometheus.io/docs/practices/naming/#base-units
@@ -174,7 +197,7 @@ def new_histogram(
         /,
         *,
         unit: Unit | None = None,
-        buckets: list[float] | None = None,
+        buckets: typing.Sequence[float] | None = None,
     ) -> Histogram:
         hist = Histogram(self, name, desc, unit, buckets=buckets)
         self._add_metric(hist)
@@ -187,7 +210,7 @@ def new_labeled_histogram(
         /,
         *,
         unit: Unit | None = None,
-        buckets: list[float] | None = None,
+        buckets: typing.Sequence[float] | None = None,
         labels: tuple[str, ...],
     ) -> LabeledHistogram:
         hist = LabeledHistogram(
@@ -505,7 +528,7 @@ class BaseHistogram(BaseMetric):
     ]
 
     def __init__(
-        self, *args: typing.Any, buckets: list[float] | None = None
+        self, *args: typing.Any, buckets: typing.Sequence[float] | None = None
     ) -> None:
         if buckets is None:
             buckets = self.DEFAULT_BUCKETS
@@ -530,7 +553,7 @@ class Histogram(BaseHistogram):
     _sum: float
 
     def __init__(
-        self, *args: typing.Any, buckets: list[float] | None = None
+        self, *args: typing.Any, buckets: typing.Sequence[float] | None = None
     ) -> None:
         super().__init__(*args, buckets=buckets)
         self._sum = 0.0
@@ -581,7 +604,7 @@ class LabeledHistogram(BaseHistogram):
     def __init__(
         self,
         *args: typing.Any,
-        buckets: list[float] | None = None,
+        buckets: typing.Sequence[float] | None = None,
         labels: tuple[str, ...],
     ) -> None:
         super().__init__(*args, buckets=buckets)

edb/server/metrics.py

@@ -24,6 +24,13 @@
 
 registry = prom.Registry(prefix='edgedb_server')
 
+COUNT_BUCKETS = prom.per_order_buckets(
+    1, 10000, entries_per_order=2,
+)
+BYTES_BUCKETS = prom.per_order_buckets(
+    32, 2**20, entries_per_order=1, base=2,
+)
+
 compiler_process_spawns = registry.new_counter(
     'compiler_process_spawns_total',
     'Total number of compiler processes spawned.'
@@ -144,13 +151,15 @@
 queries_per_connection = registry.new_labeled_histogram(
     'queries_per_connection',
     'Number of queries per connection.',
+    buckets=COUNT_BUCKETS,
     labels=('tenant', 'interface'),
 )
 
 query_size = registry.new_labeled_histogram(
     'query_size',
     'The size of a query.',
     unit=prom.Unit.BYTES,
+    buckets=BYTES_BUCKETS,
     labels=('tenant', 'interface'),
 )