feat: dfly_bench implementation and tests

README.md

@@ -56,6 +56,7 @@ examples.
 | [`setup`](#setup) | Provision an isolated VPC + EC2 fleet and pre-install every engine. | mutates |
 | [`run`](#run) | Execute a `memtier_benchmark` pass (N trials, all engines) against an existing setup. | mutates |
 | [`run-commands`](#run-commands) | Benchmark one Redis command (SET, GET, LPUSH, ...) per invocation against an existing setup. | mutates |
+| [`run-commands-dfly`](#run-commands-dfly) | Same as `run-commands`, but drives load with `dfly_bench` instead of memtier. | mutates |
 | [`run-replication`](#run-replication) | Execute the Dragonfly master/replica sweep against a setup with `--replicas >= 1`. | mutates |
 | [`quickstart`](#quickstart) | One-shot: `setup` + `run` + `destroy` in a single invocation. | mutates |
 | [`compare`](#compare) | Diff two result files (auto-detects memtier vs replication) and exit non-zero on regression. | read-only |
@@ -623,6 +624,54 @@ benchmark. 100M x 256B assumes a ~128GB server; on smaller instances
 lower it proportionally or the engines will evict or OOM during
 preload.
 
+### `run-commands-dfly`
+
+The `dfly_bench` counterpart to `run-commands`: it drives the same
+per-command flush/preload/warmup/trials cycle, but with Dragonfly's own
+load generator instead of memtier. `dfly_bench` is built from the
+`dragonflydb/dragonfly` source and installed on the client at `dfbench
+setup` time (alongside memtier); override the build with
+`--dfly-bench-ref`. Results use the same schema as `run-commands` (tagged
+`tool="dfly_bench"`) and land in `~/.dfbench/runs/<id>/commands/<timestamp>/`,
+so a memtier file and a `dfly_bench` file for the same command diff with
+`dfbench compare`.
+
+Because `dfly_bench` uses its own key prefix (`--key_prefix`, default
+`key:`), read commands (GET, ...) are preloaded by `dfly_bench` itself
+(sequential distribution, one connection per thread covering the whole
+`--key-maximum` range) rather than reusing a memtier preload.
+
+| Flag | Default | Notes |
+| ---- | ------- | ----- |
+| `--id` | (required) | Setup previously created by `dfbench setup`. |
+| `--command` | (required) | Single command name, e.g. `SET`, `GET`, `LRANGE`. |
+| `--engines` | `dragonfly` | Subset of the setup's installed engines (`dfly_bench` speaks RESP, so Redis/Valkey work too). |
+| `--dfly-threads` | `32` | `dfly_bench --proactor_threads` (analogous to memtier threads). Total connections = threads * conns. |
+| `--dfly-conns` | `5` | `dfly_bench -c`, connections per thread (analogous to memtier clients). |
+| `--dfly-pipeline` | `20` | `dfly_bench --pipeline`: max pending requests per connection. |
+| `--dfly-data-size` | `256` | `dfly_bench -d`: value size in bytes (`__data__` placeholders). |
+| `--key-maximum` | `100000000` | Key range (`--key_maximum`). Structured preloads use a built-in 1M-key range. |
+| `--key-dist` | `U` | `dfly_bench --key_dist` for the measured pass: `U` (uniform, the memtier random `R` equivalent), `N`, `Z`, `S`. |
+| `--qps` | `0` | `dfly_bench --qps` per connection; `0` = closed-loop max load (coordinated omission). |
+| `--test-time` | `300` | Seconds of measured load per trial. |
+| `--warmup-time` | `10` | Seconds of discarded warmup load per engine. `0` disables. |
+| `--trials` | `3` | Measured trials per engine (median with [min..max]). |
+| `--dfly-bench-ref` | `main` | Git ref of `dragonflydb/dragonfly` to build `dfly_bench` from if it isn't already installed. |
+| `--dragonfly_num_shards`, `--dragonfly_conn_use_incoming_cpu`, `--dragonfly_version` | (as in `run-commands`) | Server-side Dragonfly tuning; identical semantics to `run-commands`. |
+| `--redis_io_threads`, `--valkey_io_threads`, `--redis_pin_network`, `--valkey_pin_network` | (as in `run-commands`) | Redis/Valkey tuning. |
+| `--output` | `table` | `table` / `json` / `csv`. |
+
+`dfly_bench`'s `--json_out_file` writes a memtier-compatible latency
+report, so the same parser produces the throughput / p50 / p99 / p99.9 /
+avg columns for both tools.
+
+```bash
+dfbench run-commands-dfly --id percmd \
+  --engines dragonfly \
+  --command GET \
+  --key-maximum 1000000 --test-time 15 --trials 1
+```
+
 ### `run-replication`
 
 Execute the Dragonfly master/replica benchmark scenario against an
@@ -1106,7 +1155,7 @@ their keyspace can empty mid-trial. A `results.csv` with one row per
 | `run-replication` errors: `ERR Out of memory` during preload | Default `--preload-keys` and `--value-size` exceed the instance's memory budget. | Set `--master-flags "--maxmemory=25gb"` (or larger), lower `--preload-keys`, or shrink `--value-size`. Dragonfly will evict rather than OOM. |
 | Maxload case appears to run at ~20 req/sec | `dfly_bench --qps` defaults to 20 when omitted. dfbench explicitly emits `--qps=0` for `maxload`; if you see this, verify `--dfly-ref` resolves to a revision whose `dfly_bench` treats 0 as unbounded. | See `BuildDflyBenchArgs` in [`bench/replication/dflybench.go`](bench/replication/dflybench.go). |
 | `run` refuses to start with `setup "foo" is not ready (status="benchmarking")` | Previous run crashed or was killed before the deferred status-restore ran. | Rare now; `run` and `run-replication` wrap their context with `signal.NotifyContext(Interrupt, SIGTERM)` so `Ctrl+C` unwinds cleanly. If it happens: `dfbench destroy --id ...` and start over. |
-| `dfly_bench` build fails on the client: missing `helio/blaze.sh` | Submodule population failed. | `InstallDflyBench` falls back to a full clone + `git submodule update --init --recursive`. If it still fails, check the error output for missing apt packages (`bison`, `libboost-fiber-dev`, `libunwind-dev`). |
+| `dfly_bench` build fails on the client: missing `helio/blaze.sh` | Submodule population failed. | `InstallDflyBench` in [`bench/dflybench.go`](bench/dflybench.go) falls back to a full clone + `git submodule update --init --recursive`. If it still fails, check the error output for missing apt packages (`bison`, `libboost-fiber-dev`, `libunwind-dev`). |
 | `destroy --id foo` returns `setup "foo" not found` | Typo, or state file moved / deleted. | `dfbench list` to see what IDs exist. Old versions leaked the internal path; the current message is deliberately terse. |
 
 ---

bench/commands/results.go

@@ -19,7 +19,11 @@ const ScenarioTag = "commands"
 // CommandRunResults is the results.json schema for one dfbench
 // run-commands invocation.
 type CommandRunResults struct {
-	Scenario       string          `json:"scenario"` // always "commands"
+	Scenario string `json:"scenario"` // always "commands"
+	// Tool is the load generator that produced these numbers: "memtier"
+	// (dfbench run-commands) or "dfly_bench" (dfbench run-commands-dfly).
+	// It lets 'dfbench compare' diff a memtier file against a dfly_bench one.
+	Tool           string          `json:"tool,omitempty"`
 	RunID          string          `json:"run_id"`
 	SetupID        string          `json:"setup_id"`
 	CreatedAt      time.Time       `json:"created_at"`

bench/commands/runner.go

@@ -60,6 +60,20 @@ type Config struct {
 	// dfbench setup installed. Redis/Valkey ignore it.
 	DflyVersion string
 
+	// The fields below are only read by RunDfly (dfbench run-commands-dfly).
+	// The memtier Run ignores them.
+
+	// KeyDist is dfly_bench --key_dist: U (uniform, the memtier "R"
+	// equivalent), N (normal), Z (zipfian), S (sequential). Empty defaults
+	// to U for the measured pass.
+	KeyDist string
+	// QPS is dfly_bench --qps, a per-connection rate cap; 0 means closed-loop
+	// max load (coordinated omission).
+	QPS int
+	// DflyBenchRef is the dragonflydb/dragonfly git ref used to (re)build
+	// dfly_bench on the client if it isn't already installed.
+	DflyBenchRef string
+
 	OutputDir string
 }
 
@@ -102,6 +116,7 @@ func Run(
 
 	run := &CommandRunResults{
 		Scenario:       ScenarioTag,
+		Tool:           "memtier",
 		RunID:          time.Now().UTC().Format("20060102-150405"),
 		SetupID:        state.ID,
 		CreatedAt:      time.Now().UTC(),