service-capacity-modeling 0.3.54__tar.gz → 0.3.56__tar.gz

{service_capacity_modeling-0.3.54 → service_capacity_modeling-0.3.56}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: service-capacity-modeling
-Version: 0.3.54
+Version: 0.3.56
 Summary: Contains utilities for modeling capacity for pluggable workloads
 Author: Joseph Lynch
 Author-email: josephl@netflix.com

{service_capacity_modeling-0.3.54 → service_capacity_modeling-0.3.56}/service_capacity_modeling/models/org/netflix/cassandra.py

@@ -144,6 +144,7 @@ def _estimate_cassandra_requirement(  # pylint: disable=too-many-positional-argu
     disk_buffer = buffer_for_components(
         buffers=desires.buffers, components=[BufferComponent.disk]
     )
+    memory_preserve = False
     reference_shape = desires.reference_shape
     current_capacity = (
         None
@@ -169,6 +170,9 @@ def _estimate_cassandra_requirement(  # pylint: disable=too-many-positional-argu
             * current_capacity.cluster_instance_count.mid
             * (disk_scale or 1)
         )
+        _, memory_preserve = derived_buffer_for_component(
+            desires.buffers.derived, ["storage", "memory"]
+        )
     else:
         # If the cluster is not yet provisioned
         capacity_requirement = _zonal_requirement_for_new_cluster(
@@ -202,13 +206,19 @@ def _estimate_cassandra_requirement(  # pylint: disable=too-many-positional-argu
     )
     rps_working_set = min(1.0, disk_rps / max_rps_to_disk)
 
-    # If disk RPS will be smaller than our target because there are no
-    # reads, we don't need to hold as much data in memory.
-    # For c*, we can skip memory buffer and can just keep using the heap and write buffer calc
-    # Eventually we'll want to phrase those heap, read cache, and write cache as buffers
-    needed_memory = min(working_set, rps_working_set) * disk_used_gib * zones_per_region
-    # Now convert to per zone
-    needed_memory = max(1, int(needed_memory // zones_per_region))
+    if memory_preserve:
+        needed_memory = capacity_requirement.mem_gib.mid
+    else:
+        # If disk RPS will be smaller than our target because there are no
+        # reads, we don't need to hold as much data in memory.
+        # For c*, we can skip memory buffer and can just keep using the heap and write buffer calc
+        # Eventually we'll want to phrase those heap, read cache, and write cache as buffers
+        needed_memory = (
+            min(working_set, rps_working_set) * disk_used_gib * zones_per_region
+        )
+        # Now convert to per zone
+        needed_memory = max(1, int(needed_memory // zones_per_region))
+
     logger.debug(
         "Need (cpu, mem, disk, working) = (%s, %s, %s, %f)",
         needed_cores,
@@ -381,6 +391,7 @@ def _estimate_cassandra_cluster_zonal(  # pylint: disable=too-many-positional-ar
         # C* clusters provision in powers of 2 because doubling
         cluster_size=next_power_of_2,
         min_count=max(min_count, required_cluster_size or 0),
+        # TODO: Take reserve memory calculation into account during buffer calculation
         # C* heap usage takes away from OS page cache memory
         reserve_memory=lambda x: base_mem + heap_fn(x),
         # C* heap buffers the writes at roughly a rate of

{service_capacity_modeling-0.3.54 → service_capacity_modeling-0.3.56}/service_capacity_modeling/models/org/netflix/kafka.py

@@ -188,7 +188,8 @@ def _estimate_kafka_requirement(  # pylint: disable=too-many-positional-argument
     ) // zones_per_region
 
     logger.debug(
-        "Need (cpu, mem, disk) = (%s, %s, %s)",
+        "Need (instance, cpu, mem, disk) = (%s, %s, %s, %s)",
+        instance.name,
         needed_cores,
         needed_memory,
         needed_disk,
@@ -235,7 +236,7 @@ def _kafka_read_io(rps, io_size_kib, size_gib, recovery_seconds: int) -> float:
 # pylint: disable=too-many-locals
 # pylint: disable=too-many-return-statements
 # pylint: disable=too-many-positional-arguments
-def _estimate_kafka_cluster_zonal(
+def _estimate_kafka_cluster_zonal(  # noqa: C901
     instance: Instance,
     drive: Drive,
     desires: CapacityDesires,
@@ -360,11 +361,14 @@ def _estimate_kafka_cluster_zonal(
     if required_zone_size is not None:
         space_gib = max(1, math.ceil(requirement.disk_gib.mid / required_zone_size))
         ebs_gib = utils.next_n(space_gib, n=100)
-        max_size = (
-            max_attached_disk_gib
-            if max_attached_disk_gib is not None
-            else drive.max_size_gib / 3
-        )  # Max allowed disk in `compute_stateful_zone`
+
+        # Max allowed disk size in `compute_stateful_zone`
+        if instance.drive is not None and instance.drive.size_gib > 0:
+            max_size = min(max_local_disk_gib, instance.drive.size_gib)
+        elif max_attached_disk_gib is not None:
+            max_size = max_attached_disk_gib
+        else:
+            max_size = drive.max_size_gib / 3
 
         # Capacity planner only allows ~ 5TB disk (max_size) for gp3 drives
         # or max_attached_disk_gib if provided.

{service_capacity_modeling-0.3.54 → service_capacity_modeling-0.3.56}/service_capacity_modeling.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: service-capacity-modeling
-Version: 0.3.54
+Version: 0.3.56
 Summary: Contains utilities for modeling capacity for pluggable workloads
 Author: Joseph Lynch
 Author-email: josephl@netflix.com