diff --git a/Cargo.toml b/Cargo.toml
index 6487165b..0be6323b 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -1,5 +1,6 @@
 [workspace]
 members = [
+   "async_data_factory",
    "async_db",
    "async_ex1",
    "async_ex2",
diff --git a/async_data_factory/Cargo.toml b/async_data_factory/Cargo.toml
new file mode 100644
index 00000000..47f6d73c
--- /dev/null
+++ b/async_data_factory/Cargo.toml
@@ -0,0 +1,15 @@
+[package]
+name = "async_data_factory"
+version = "0.1.0"
+authors = ["fakeshadow <24548779@qq.com>"]
+edition = "2018"
+workspace = ".."
+
+
+[dependencies]
+actix-rt = "1.1.1"
+actix-web = { version = "2.0.0" }
+num_cpus = "1.13.0"
+redis = { version = "0.16.0", default-features = false, features = ["tokio-rt-core"] }
+# redis_tang is an redis pool for test purpose
+redis_tang = "0.1.0"
diff --git a/async_data_factory/README.md b/async_data_factory/README.md
new file mode 100644
index 00000000..98c31e07
--- /dev/null
+++ b/async_data_factory/README.md
@@ -0,0 +1,23 @@
+## Usage:
+This is an example on constructing async state with `App::data_factory`
+
+## Reason:
+`data_factory` would make sense in these situations:
+- When async state not necessarily have to be shared between workers/threads.
+
+- When async state would spawn tasks on `actix-rt`. If we centralized the state there could be a possibility the tasks get a very unbalanced distribution on the workers/threads
+(`actix-rt` would spawn tasks on local thread whenever it's called)
+
+## Requirement:
+- `rustc 1.43 stable`
+- `redis` server listen on `127.0.0.1:6379`(or use `REDIS_URL` env argument when starting the example)
+
+## Endpoints:
+- use a work load generator(e.g wrk) to benchmark the end points:
+
+        http://127.0.0.1:8080/pool   prebuilt shared redis pool
+        http://127.0.0.1:8080/pool2  data_factory redis pool
+
+## Context:
+The real world difference can be vary by the work you are doing but in general it's a good idea to
+spread your *identical* async tasks evenly between threads and have as little cross threads synchronization as possible.
diff --git a/async_data_factory/src/main.rs b/async_data_factory/src/main.rs
new file mode 100644
index 00000000..007af704
--- /dev/null
+++ b/async_data_factory/src/main.rs
@@ -0,0 +1,106 @@
+use actix_web::web::Data;
+use actix_web::{get, App, HttpServer};
+
+use redis_tang::{Builder, Pool, RedisManager};
+
+#[actix_rt::main]
+async fn main() -> std::io::Result<()> {
+    let redis_url =
+        std::env::var("REDIS_URL").unwrap_or_else(|_| String::from("redis://127.0.0.1"));
+
+    let num_cpus = num_cpus::get();
+
+    // a shared redis pool for work load comparison.
+    let pool = pool_builder(num_cpus, redis_url.as_str())
+        .await
+        .expect("fail to build pool");
+    let pool = RedisWrapper(pool);
+
+    HttpServer::new(move || {
+        let redis_url = redis_url.clone();
+
+        App::new()
+            .data(pool.clone())
+            // a dummy data_factory implementation
+            .data_factory(|| {
+                /*
+                    App::data_factory would accept a future as return type and poll the future when
+                    App is initialized.
+
+                    The Output of the future must be Result<T, E> and T will be the transformed to
+                    App::Data<T> that can be extracted from handler/request.
+                    (The E will only be used to trigger a log::error.)
+
+                    This data is bound to worker thread and you get an instance of it for every
+                    worker of the HttpServer.(hence the name data_factory)
+                    *. It is NOT shared between workers
+                    (unless the underlying data is a smart pointer like Arc<T>).
+                */
+
+                async {
+                    // 123usize would be transformed into Data<usize>
+                    Ok::<usize, ()>(123)
+                }
+            })
+            // a data_factory redis pool for work load comparison.
+            .data_factory(move || pool_builder(1, redis_url.clone()))
+            .service(pool_shared_prebuilt)
+            .service(pool_local)
+    })
+    .bind("127.0.0.1:8080")?
+    .run()
+    .await
+}
+
+/*
+    This pool is shared between workers. We have all redis connections spawned tasks on main thread
+    therefore it puts too much pressure on one thread.
+    *. This is the case for redis::aio::MultiplexedConnection and it may not apply to other async
+    redis connection type.
+*/
+#[get("/pool")]
+async fn pool_shared_prebuilt(pool: Data<RedisWrapper>) -> &'static str {
+    ping(&pool.as_ref().0).await
+}
+
+/*
+   This pool is built with App::data_factory and we have 2 connections fixed for every worker.
+   It's evenly distributed and have no cross workers synchronization.
+*/
+#[get("/pool2")]
+async fn pool_local(data: Data<usize>, pool: Data<Pool<RedisManager>>) -> &'static str {
+    assert_eq!(data.get_ref(), &123);
+
+    ping(pool.as_ref()).await
+}
+
+// boiler plate for redis pool
+#[derive(Clone)]
+struct RedisWrapper(Pool<RedisManager>);
+
+async fn pool_builder(
+    num_cpus: usize,
+    redis_url: impl redis::IntoConnectionInfo,
+) -> Result<Pool<RedisManager>, ()> {
+    let mgr = RedisManager::new(redis_url);
+    Builder::new()
+        .always_check(false)
+        .idle_timeout(None)
+        .max_lifetime(None)
+        .min_idle(num_cpus * 2)
+        .max_size(num_cpus * 2)
+        .build(mgr)
+        .await
+        .map_err(|_| ())
+}
+
+async fn ping(pool: &Pool<RedisManager>) -> &'static str {
+    let mut client = pool.get().await.unwrap().clone();
+
+    redis::cmd("PING")
+        .query_async::<_, ()>(&mut client)
+        .await
+        .unwrap();
+
+    "Done"
+}