compute: pack accumulable reduce accumulators at their natural width

misc/python/materialize/mzcompose/__init__.py

@@ -271,6 +271,11 @@ def get_variable_system_parameters(
         VariableSystemParameter(
             "enable_columnar_merge_batcher", "true", ["true", "false"]
         ),
+        # On by default so CI exercises the packed accumulable diff layout, which
+        # is off in production while it earns trust.
+        VariableSystemParameter(
+            "enable_packed_accumulable_diff", "true", ["true", "false"]
+        ),
         VariableSystemParameter(
             "compute_peek_response_stash_threshold_bytes",
             # 1 MiB, an in-between value

misc/python/materialize/parallel_workload/action.py

@@ -3104,6 +3104,7 @@ def __init__(
         self.flags_with_values["enable_column_paged_batcher_spill"] = (
             BOOLEAN_FLAG_VALUES
         )
+        self.flags_with_values["enable_packed_accumulable_diff"] = BOOLEAN_FLAG_VALUES
         # Fractions of the *cgroup* memory limit, which under mzcompose's
         # process orchestrator is the whole container budget shared by
         # environmentd and every replica, not one replica's allowance. A

src/compute-types/src/dyncfgs.rs

@@ -44,6 +44,20 @@ pub const ENABLE_ERROR_DISTINCT: Config<bool> = Config::new(
     ParameterScope::Environment,
 );
 
+/// Select the packed byte layout for the accumulable reduce's arrangement diff.
+///
+/// The accumulators are stored at their natural widths instead of at the size of
+/// the widest enum variant, so a group's diff shrinks to the sum of its slot
+/// sizes. Read at operator construction time, so flips take effect on dataflows
+/// created after the change.
+pub const ENABLE_PACKED_ACCUMULABLE_DIFF: Config<bool> = Config::new(
+    "enable_packed_accumulable_diff",
+    false,
+    "Select the packed byte layout for the accumulable reduce's arrangement diff, instead of a \
+     vector of fixed-size accumulator enums.",
+    ParameterScope::Replica,
+);
+
 /// Use the column-paged merge batcher code path at arrange sites. When
 /// `true`, arrange operators use `Col2ValPagedBatcher` (in
 /// `mz_timely_util::columnar`) and `RowRowColPagedBuilder` (in
@@ -702,6 +716,7 @@ pub fn all_dyncfgs(configs: ConfigSet) -> ConfigSet {
     configs
         .add(&ENABLE_HALF_JOIN2)
         .add(&ENABLE_ERROR_DISTINCT)
+        .add(&ENABLE_PACKED_ACCUMULABLE_DIFF)
         .add(&ENABLE_MZ_JOIN_CORE)
         .add(&ENABLE_SYNC_MV_SINK)
         .add(&ENABLE_CORRECTION_V2)