Adds support for disk_io in evcache

service_capacity_modeling/models/common.py

@@ -199,6 +199,7 @@ def compute_stateless_region(
 
 
 # pylint: disable=too-many-locals
+# pylint: disable=too-many-statements
 def compute_stateful_zone(
     instance: Instance,
     drive: Drive,
@@ -229,6 +230,8 @@ def compute_stateful_zone(
     # Some stateful clusters have preferences on per zone sizing
     cluster_size: Callable[[int], int] = lambda x: x,
     min_count: int = 0,
+    adjusted_disk_io_needed: float = 0.0,
+    read_write_ratio: float = 0.0,
 ) -> ZoneClusterCapacity:
     # Normalize the cores of this instance type to the latency reference
     needed_cores = math.ceil(
@@ -269,6 +272,31 @@ def compute_stateful_zone(
     ):
         disk_per_node = min(max_local_disk_gib, instance.drive.size_gib)
         count = max(count, math.ceil(needed_disk_gib / disk_per_node))
+        if adjusted_disk_io_needed != 0.0:
+            instance_read_iops = (
+                instance.drive.read_io_per_s
+                if instance.drive.read_io_per_s is not None
+                else 0
+            )
+            assert isinstance(instance_read_iops, int)
+            instance_write_iops = (
+                instance.drive.write_io_per_s
+                if instance.drive.write_io_per_s is not None
+                else 0
+            )
+            assert isinstance(instance_write_iops, int)
+            instance_adjusted_io = (
+                (
+                    read_write_ratio * float(instance_read_iops)
+                    + (1.0 - read_write_ratio) * float(instance_write_iops)
+                )
+                * instance.drive.block_size_kib
+                * 1024.0
+            )
+            if instance_adjusted_io != 0.0:
+                count = max(
+                    count, math.ceil(adjusted_disk_io_needed / instance_adjusted_io)
+                )
 
     count = max(cluster_size(count), min_count)
     cost = count * instance.annual_cost

service_capacity_modeling/models/org/netflix/evcache.py

@@ -232,6 +232,21 @@ def reserve_memory(instance_mem_gib):
     if desires.service_tier < 1:
         min_count = 2
 
+    is_disk_io_constraint: bool = requirement.disk_gib.mid > 0.0
+    adjusted_disk_io_needed = 0.0
+    read_write_ratio = 0.0
+    if is_disk_io_constraint:
+        reads_per_sec = desires.query_pattern.estimated_read_per_second.mid
+        writes_per_sec = desires.query_pattern.estimated_write_per_second.mid
+        read_size = desires.query_pattern.estimated_mean_read_size_bytes.mid
+        write_size = desires.query_pattern.estimated_mean_write_size_bytes.mid
+        read_disk_io_needed = reads_per_sec * read_size
+        write_disk_io_needed = writes_per_sec * write_size
+        adjusted_disk_io_needed = read_disk_io_needed + write_disk_io_needed
+        # Giving headroom for cachewarming and region squeeze
+        adjusted_disk_io_needed = 1.4 * adjusted_disk_io_needed
+        read_write_ratio = reads_per_sec / (reads_per_sec + writes_per_sec)
+
     cluster = compute_stateful_zone(
         instance=instance,
         drive=drive,
@@ -248,6 +263,8 @@ def reserve_memory(instance_mem_gib):
         # Sidecars and Variable OS Memory
         reserve_memory=lambda x: base_mem,
         core_reference_ghz=requirement.core_reference_ghz,
+        adjusted_disk_io_needed=adjusted_disk_io_needed,
+        read_write_ratio=read_write_ratio,
     )
     # Communicate to the actual provision that if we want reduced RF
     params = {"evcache.copies": copies_per_region}