ci: improve bench by data size

.github/workflows/build_test_linux.yml

@@ -10,6 +10,30 @@ env:
   RUST_BACKTRACE: full
 
 jobs:
+  build_bench:
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: true
+      matrix:
+        os: [ubuntu-latest]
+        toolchain: [stable]
+    steps:
+      - uses: actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11 # v4
+      - name: Setup Rust
+        uses: actions-rs/toolchain@v1
+        with:
+          toolchain: ${{ matrix.toolchain }}
+          override: true
+      - uses: extractions/setup-just@v1
+      - uses: hustcer/setup-nu@v3.8
+        with:
+          version: '0.85'
+        env:
+            GITHUB_TOKEN: ${{ secrets.PAT_GLOBAL }}
+      - name: Just version
+        run: just --version
+      - name: Build
+        run: just bench_build
   build_test:
     runs-on: ${{ matrix.os }}
     strategy:

Cargo.toml

@@ -40,7 +40,8 @@ criterion = { version = "0.5.1" }
 doc-comment = "0.3.3"
 uuid = { version = "1", features = ["serde", "v4"] }
 chrono = { version = "0.4", features = ["serde"] }
-
+rand = "0.8"
+once_cell = "1.19"
 
 [features]
 default = []

README.md

@@ -22,7 +22,7 @@ Here's a drop-in, fast, embedded database for multi-platform apps (server, deskt
 
 - Simple API 🦀.
 - Support for **multiple indexes** (primary, secondary, unique, non-unique, optional).
-- Minimal boilerplate.
+- Minimal boilerplate see benchmarks [here](./benches).
 - Transparent serialization/deserialization using [native_model](https://github.com/vincent-herlemont/native_model).
 - **Automatic model migration** 🌟.
 - **Thread-safe** and fully **ACID-compliant** transactions provided by [redb](https://github.com/cberner/redb).

benches/README.md

@@ -0,0 +1,24 @@
+# Benchmarks
+
+Highlight the Native DB overhead compared to direct access to the underlying database ([redb](https://github.com/cberner/redb)).
+
+The benchmarks ignore:
+- [`native_model`](https://github.com/vincent-herlemont/native_model) overhead.
+- Serialization overhead used by `native_model` like `bincode`,`postcard` etc.
+- The fact that `redb` can copy the data using zero-copy.
+
+## Run benchmarks
+
+If you want to run the benchmarks, you need to install [just](https://github.com/casey/just), [nushell](https://www.nushell.sh/) and run:
+```bash
+just bench
+```
+
+## Results
+
+We can see that the overhead is very low. These result are obtained with the version `0.5.3` of `native_db`.
+
+![](./results/insert_random.png)
+![](./results/get_random.png)
+![](./results/remove_random.png)
+![](./results/scan_random.png)

benches/overhead_data_size.rs

@@ -1,8 +1,48 @@
 use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion};
 use native_db::*;
 use native_model::{native_model, Model};
-use redb::TableDefinition;
+use redb::{ReadableTable, TableDefinition};
 use serde::{Deserialize, Serialize};
+use once_cell::sync::Lazy;
+use rand::prelude::SliceRandom;
+
+// 1 byte * 10000, 10 bytes * 10000, 100 bytes * 5000, 1KB * 1000, 1MB * 100, 10MB * 10
+const ITERATIONS:&'static [(usize, usize)] = &[(1, 10000), (10, 10000), (100, 5000), (1024, 1000), (1024 * 1024, 100), (10 * 1024 * 1024, 10)];
+
+static DATABASE_BUILDER: Lazy<DatabaseBuilder> = Lazy::new(|| {
+    let mut builder = DatabaseBuilder::new();
+    builder.define::<Data>().unwrap();
+    builder
+});
+
+fn init_database() -> (redb::Database, Database<'static>) {
+    let redb_backend = redb::backends::InMemoryBackend::new();
+    let redb_db = redb::Database::builder()
+        .create_with_backend(redb_backend)
+        .unwrap();
+
+    let native_db = DATABASE_BUILDER.create_in_memory().unwrap();
+    (redb_db, native_db)
+}
+
+fn generate_random_data(redb_db: &redb::Database, native_db: &Database,nb_bytes: &usize, nb_items: &usize) -> Vec<Data> {
+    let data = Data {
+        x: 1,
+        data: vec![1u8; *nb_bytes],
+    };
+
+    let mut out = vec![];
+
+    for _ in 0..*nb_items {
+        let mut data = data.clone();
+        data.random_x();
+        use_redb_insert(&redb_db, data.clone());
+        use_native_db_insert(&native_db, data.clone());
+        out.push(data);
+    }
+
+    out
+}
 
 #[derive(Serialize, Deserialize, Clone)]
 #[native_model(id = 1, version = 1)]
@@ -13,51 +53,240 @@ struct Data {
     data: Vec<u8>,
 }
 
+impl Data {
+    fn random_x(&mut self) {
+        self.x = rand::random();
+    }
+}
+
 const TABLE_REDB: TableDefinition<u32, &'static [u8]> = TableDefinition::new("my_data");
-fn use_redb(db: &redb::Database, data: Data) {
+
+fn use_redb_insert(db: &redb::Database, data: Data) {
     let rw = db.begin_write().unwrap();
     {
         let mut table = rw.open_table(TABLE_REDB).unwrap();
+        // Because native_db use native_model to encode data, we do the same here
+        // to remove the overhead of the encoding.
         let encode = native_model::encode(&data).unwrap();
         table.insert(data.x, encode.as_slice()).unwrap();
     }
     rw.commit().unwrap();
 }
 
-fn use_native_db(db: &native_db::Database, data: Data) {
+fn use_redb_get(db: &redb::Database, x: u32) -> Data {
+    let ro = db.begin_read().unwrap();
+    let out;
+    {
+        let table = ro.open_table(TABLE_REDB).unwrap();
+        out = table.get(x).unwrap().map(|v| {
+            native_model::decode(v.value().to_vec()).unwrap().0
+        }).expect("Data not found");
+    }
+    out
+}
+
+fn use_redb_scan(db: &redb::Database) -> Vec<Data> {
+    let ro = db.begin_read().unwrap();
+    let out;
+    {
+        let table = ro.open_table(TABLE_REDB).unwrap();
+        out = table.iter().unwrap().map(|r| {
+            let (_, v) = r.unwrap();
+            native_model::decode(v.value().to_vec()).unwrap().0
+        }).collect::<Vec<Data>>();
+    }
+    out
+}
+
+fn redb_remove(db: &redb::Database, x: u32) {
+    let rw = db.begin_write().unwrap();
+    {
+        let mut table = rw.open_table(TABLE_REDB).unwrap();
+        table.remove(x).unwrap().expect("Data not found");
+    }
+    rw.commit().unwrap();
+}
+
+fn use_native_db_insert(db: &Database, data: Data) {
     let rw = db.rw_transaction().unwrap();
     rw.insert(data).unwrap();
     rw.commit().unwrap();
 }
 
-fn criterion_benchmark(c: &mut Criterion) {
-    let mut group = c.benchmark_group("insert");
+fn use_native_db_scan(db: &Database) -> Vec<Data> {
+    let r = db.r_transaction().unwrap();
+    let out = r.scan().primary().unwrap().all().collect::<Vec<_>>();
+    out
+}
+
+fn use_native_db_get(db: &Database, x: u32) -> Data {
+    let r = db.r_transaction().unwrap();
+    let out = r.get().primary(x).unwrap().unwrap();
+    out
+}
+
+fn native_db_remove(db: &Database, data: Data) {
+    let rw = db.rw_transaction().unwrap();
+    rw.remove(data).unwrap();
+    rw.commit().unwrap();
+}
+
+// Benchmarks
+
+fn bench_get_random(c: &mut Criterion) {
+    let mut group = c.benchmark_group("get_random");
+    let plot_config = criterion::PlotConfiguration::default().summary_scale(criterion::AxisScale::Logarithmic);
+    group.plot_config(plot_config.clone());
+    group.sampling_mode(criterion::SamplingMode::Flat);
+
+    for (nb_bytes,nb_items) in ITERATIONS {
+        group.throughput(criterion::Throughput::Bytes(*nb_bytes as u64));
+
+        let (redb_db, native_db) = init_database();
+        let data = generate_random_data(&redb_db, &native_db, nb_bytes, nb_items);
+
+        group.bench_function(BenchmarkId::new("redb", nb_bytes), |b| {
+            b.iter_batched(
+                || {
+                    let item = data.choose(&mut rand::thread_rng()).unwrap();
+                    item.x
+                },
+                |x| use_redb_get(&redb_db, x),
+                criterion::BatchSize::SmallInput
+            );
+        });
+        group.bench_function(BenchmarkId::new("native_db", nb_bytes), |b| {
+            b.iter_batched(
+                || {
+                    let item = data.choose(&mut rand::thread_rng()).unwrap();
+                    item.x
+                },
+                |x| use_native_db_get(&native_db, x),
+                criterion::BatchSize::SmallInput
+            );
+        });
+    }
+}
+
 
-    // 1 byte, 1KB, 1MB, 10MB, 100MB
-    for nb_bytes in [1, 1024, 1024 * 1024, 10 * 1024 * 1024, 100 * 1024 * 1024] {
-        group.throughput(criterion::Throughput::Bytes(nb_bytes as u64));
+fn bench_scan_random(c: &mut Criterion) {
+    let plot_config = criterion::PlotConfiguration::default().summary_scale(criterion::AxisScale::Logarithmic);
+    let mut group = c.benchmark_group("scan_random");
+    group.plot_config(plot_config.clone());
+    group.sampling_mode(criterion::SamplingMode::Flat);
+
+    for (nb_bytes,nb_items) in ITERATIONS {
+        group.throughput(criterion::Throughput::Bytes(*nb_bytes as u64));
+
+        let (redb_db, native_db) = init_database();
+        generate_random_data(&redb_db, &native_db, nb_bytes, nb_items);
+
+        group.bench_function(BenchmarkId::new("redb", nb_bytes), |b| {
+            b.iter_with_large_drop(|| use_redb_scan(&redb_db));
+        });
+
+        group.bench_function(BenchmarkId::new("native_db", nb_bytes), |b| {
+            b.iter_with_large_drop(|| use_native_db_scan(&native_db));
+        });
+    }
+}
+
+
+fn bench_remove_random(c: &mut Criterion) {
+    let mut group = c.benchmark_group("remove_random");
+    let plot_config = criterion::PlotConfiguration::default().summary_scale(criterion::AxisScale::Logarithmic);
+    group.plot_config(plot_config.clone());
+    group.sampling_mode(criterion::SamplingMode::Flat);
+
+    for (nb_bytes,nb_items) in ITERATIONS {
+        group.throughput(criterion::Throughput::Bytes(*nb_bytes as u64));
+
+        let (redb_db, native_db) = init_database();
+
+        group.bench_function(BenchmarkId::new("redb", nb_bytes), |b| {
+            b.iter_batched(
+                || {
+                    let mut data = Data {
+                        x: 1,
+                        data: vec![1u8; *nb_bytes as usize],
+                    };
+                    data.random_x();
+                    use_redb_insert(&redb_db, data.clone());
+                    data
+                },
+                |data| redb_remove(&redb_db, data.x),
+                criterion::BatchSize::SmallInput
+            );
+        });
+
+        group.bench_function(BenchmarkId::new("native_db", nb_bytes), |b| {
+            b.iter_batched(
+                || {
+                    let mut data = Data {
+                        x: 1,
+                        data: vec![1u8; *nb_bytes as usize],
+                    };
+                    data.random_x();
+                    use_native_db_insert(&native_db, data.clone());
+                    data
+                },
+                |data| native_db_remove(&native_db, data),
+                criterion::BatchSize::SmallInput
+            );
+        });
+    }
+}
+
+
+fn bench_insert_random(c: &mut Criterion) {
+    let mut insert_random_group = c.benchmark_group("insert_random");
+    let plot_config = criterion::PlotConfiguration::default().summary_scale(criterion::AxisScale::Logarithmic);
+    insert_random_group.plot_config(plot_config.clone());
+    insert_random_group.sampling_mode(criterion::SamplingMode::Flat);
+
+    // 1 byte, 10 bytes, 100 bytes, 1KB, 1MB, 10MB
+    for (nb_bytes,_) in ITERATIONS {
+        insert_random_group.throughput(criterion::Throughput::Bytes(*nb_bytes as u64));
 
         let data = Data {
             x: 1,
-            data: vec![1u8; nb_bytes as usize],
+            data: vec![1u8; *nb_bytes as usize],
         };
 
-        let redb_backend = redb::backends::InMemoryBackend::new();
-        let redb_db = redb::Database::builder()
-            .create_with_backend(redb_backend)
-            .unwrap();
+        let (redb_db, native_db) = init_database();
 
-        group.bench_function(BenchmarkId::new("redb", nb_bytes), |b| {
-            b.iter(|| use_redb(&redb_db, data.clone()))
+        let batch_size = match nb_bytes {
+            nb_bytes if *nb_bytes < 1024 => criterion::BatchSize::SmallInput,
+            nb_bytes if *nb_bytes < 1024 * 1024 => criterion::BatchSize::LargeInput,
+            _ => criterion::BatchSize::PerIteration,
+        };
+
+        insert_random_group.bench_function(BenchmarkId::new("redb", nb_bytes), |b| {
+            b.iter_batched(
+                || {
+                    let mut data = data.clone();
+                    data.random_x();
+                    data
+                },
+                |data| use_redb_insert(&redb_db, data),
+                batch_size
+            );
         });
 
-        let mut native_db = native_db::Database::create_in_memory().unwrap();
-        native_db.define::<Data>().unwrap();
-        group.bench_function(BenchmarkId::new("native_db", nb_bytes), |b| {
-            b.iter(|| use_native_db(&native_db, data.clone()))
+        insert_random_group.bench_function(BenchmarkId::new("native_db", nb_bytes), |b| {
+            b.iter_batched(
+                || {
+                    let mut data = data.clone();
+                    data.random_x();
+                    data
+                },
+                |data| use_native_db_insert(&native_db, data),
+                batch_size
+            );
         });
     }
+    insert_random_group.finish();
 }
 
-criterion_group!(benches, criterion_benchmark);
+criterion_group!(benches, bench_insert_random, bench_scan_random, bench_get_random, bench_remove_random);
 criterion_main!(benches);