mz_storage/upsert/
autospill.rs

// Copyright Materialize, Inc. and contributors. All rights reserved.
//
// Use of this software is governed by the Business Source License
// included in the LICENSE file.
//
// As of the Change Date specified in that file, in accordance with
// the Business Source License, use of this software will be governed
// by the Apache License, Version 2.0.

//! An `UpsertStateBackend` that starts in memory and spills to RocksDB
//! when the total size passes some threshold.

use std::sync::Arc;

use mz_ore::metrics::DeleteOnDropGauge;
use prometheus::core::AtomicU64;
use serde::{de::DeserializeOwned, Serialize};

use super::memory::InMemoryHashMap;
use super::rocksdb::RocksDB;
use super::types::{
    GetStats, MergeStats, MergeValue, PutStats, PutValue, StateValue, UpsertStateBackend,
    UpsertValueAndSize,
};
use super::UpsertKey;

pub enum BackendType<O> {
    InMemory(InMemoryHashMap<O>),
    RocksDb(RocksDB<O>),
}

pub struct AutoSpillBackend<O, F> {
    backend_type: BackendType<O>,
    auto_spill_threshold_bytes: usize,
    rocksdb_autospill_in_use: Arc<DeleteOnDropGauge<'static, AtomicU64, Vec<String>>>,
    rocksdb_init_fn: Option<F>,
}

impl<O, F, Fut> AutoSpillBackend<O, F>
where
    O: Clone + Send + Sync + Serialize + DeserializeOwned + 'static,
    F: FnOnce() -> Fut + 'static,
    Fut: std::future::Future<Output = RocksDB<O>>,
{
    pub(crate) fn new(
        rocksdb_init_fn: F,
        auto_spill_threshold_bytes: usize,
        rocksdb_autospill_in_use: Arc<DeleteOnDropGauge<'static, AtomicU64, Vec<String>>>,
    ) -> Self {
        // Initializing the metric to 0, to reflect in memory hash map is being used
        rocksdb_autospill_in_use.set(0);
        Self {
            backend_type: BackendType::InMemory(InMemoryHashMap::default()),
            rocksdb_init_fn: Some(rocksdb_init_fn),
            auto_spill_threshold_bytes,
            rocksdb_autospill_in_use,
        }
    }
}

#[async_trait::async_trait(?Send)]
impl<O, F, Fut> UpsertStateBackend<O> for AutoSpillBackend<O, F>
where
    O: Clone + Send + Sync + Serialize + DeserializeOwned + 'static,
    F: FnOnce() -> Fut + 'static,
    Fut: std::future::Future<Output = RocksDB<O>>,
{
    fn supports_merge(&self) -> bool {
        // We only support merge if the backend supports it; the in-memory backend does not
        // and the rocksdb backend does if configure to do so.
        match &self.backend_type {
            BackendType::InMemory(_) => false,
            BackendType::RocksDb(backend) => backend.supports_merge(),
        }
    }

    async fn multi_put<P>(&mut self, puts: P) -> Result<PutStats, anyhow::Error>
    where
        P: IntoIterator<Item = (UpsertKey, PutValue<StateValue<O>>)>,
    {
        // Note that we never revert back to memory if the size shrinks below the threshold.
        // That case is considered rare and not worth the complexity.
        match &mut self.backend_type {
            BackendType::InMemory(map) => {
                let mut put_stats = map.multi_put(puts).await?;
                let in_memory_size: usize = map
                    .current_size()
                    .try_into()
                    .expect("unexpected error while casting");
                if in_memory_size > self.auto_spill_threshold_bytes {
                    tracing::info!("spilling to disk for upsert");
                    let mut rocksdb_backend =
                        self.rocksdb_init_fn
                            .take()
                            .expect("Can only initialize once")()
                        .await;

                    let (last_known_size, new_puts) = map.drain();
                    let new_puts = new_puts.map(|(k, v)| {
                        (
                            k,
                            PutValue {
                                value: Some(v),
                                previous_value_metadata: None,
                            },
                        )
                    });

                    let rocksdb_stats = rocksdb_backend.multi_put(new_puts).await?;
                    // Adjusting the sizes as the value sizes in rocksdb could be different than in memory
                    put_stats.size_diff += rocksdb_stats.size_diff;
                    put_stats.size_diff -= last_known_size;
                    // Setting backend to rocksdb
                    self.backend_type = BackendType::RocksDb(rocksdb_backend);
                    // Switching metric to 1 for rocksdb
                    self.rocksdb_autospill_in_use.set(1);
                }
                Ok(put_stats)
            }
            BackendType::RocksDb(rocks_db) => rocks_db.multi_put(puts).await,
        }
    }

    async fn multi_merge<M>(&mut self, merges: M) -> Result<MergeStats, anyhow::Error>
    where
        M: IntoIterator<Item = (UpsertKey, MergeValue<StateValue<O>>)>,
    {
        match &mut self.backend_type {
            BackendType::InMemory(_) => {
                anyhow::bail!("InMemoryHashMap does not support merging");
            }
            BackendType::RocksDb(rocks_db) => rocks_db.multi_merge(merges).await,
        }
    }

    async fn multi_get<'r, G, R>(
        &mut self,
        gets: G,
        results_out: R,
    ) -> Result<GetStats, anyhow::Error>
    where
        G: IntoIterator<Item = UpsertKey>,
        R: IntoIterator<Item = &'r mut UpsertValueAndSize<O>>,
        O: 'r,
    {
        match &mut self.backend_type {
            BackendType::InMemory(in_memory_hash_map) => {
                in_memory_hash_map.multi_get(gets, results_out).await
            }
            BackendType::RocksDb(rocks_db) => rocks_db.multi_get(gets, results_out).await,
        }
    }
}
mz_storage/upsert/autospill.rs

mz_storage/upsert/
autospill.rs
