//! [`ManagedDb`] and [`StateSyncDb`] implementations for QMDB //! [`any`](commonware_storage::qmdb::any) databases. //! //! Batch reads consult the database at the time of the read for keys their batch chain does not //! write, and return a `StaleRead` error once the database is at no state of that chain. use crate::stateful::db::{ BatchContext, InitError, ManagedDb, Merkleized as MerkleizedTrait, Shared, StateSyncDb, SyncEngineConfig, Unmerkleized as UnmerkleizedTrait, sync_standard_db, validate_initialization, }; use commonware_codec::{Codec, Read as CodecRead}; use commonware_cryptography::Hasher; use commonware_parallel::Strategy; use commonware_runtime::{Handle, Spawner}; use commonware_storage::{ Context, index::{ Ordered as OrderedIndex, Unordered as UnorderedIndex, unordered::Index as UnorderedIdx, }, journal::contiguous::{ Contiguous, Mutable, fixed::Journal as FixedJournal, variable::Journal as VariableJournal, }, merkle::{Family, Location}, qmdb::{ Error, any::{ FixedConfig, VariableConfig, batch::{MerkleizedBatch, Slot, Staged, UnmerkleizedBatch}, db::Db, initial_root, operation::{Operation, Update}, ordered, unordered, value::{self, FixedEncoding, ValueEncoding, VariableEncoding}, }, floor::Proportional, operation::Key, sync::{self, Target as AnySyncTarget}, }, translator::Translator, }; use commonware_utils::{ Array, channel::{mpsc, ring}, non_empty_range, }; use std::{ ops::{Deref, Range}, sync::Arc, }; // Matches commonware_storage::qmdb::any::BITMAP_CHUNK_BYTES, which is crate-private. const ANY_BITMAP_CHUNK_BYTES: usize = 64; /// A speculative batch of updates and deletes over a shared `any` database. pub struct AnyUnmerkleized where F: Family, E: Context, U: Update, C: Contiguous>, I: UnorderedIndex>, H: Hasher, S: Strategy, Operation: Codec, { batch: UnmerkleizedBatch, db: Shared>, metadata: Option, } /// A staged batch returned by [`AnyUnmerkleized::stage`]. /// /// Like any speculative batch, this handle reads the shared database only while the database is /// at a state of the batch's chain, and its reads return [`Error::StaleRead`] once the database /// is at any other state (see [`MerkleizedBatch`]'s "Batch validity" docs). pub struct AnyStaged where F: Family, E: Context, U: Update, C: Contiguous>, I: UnorderedIndex>, H: Hasher, S: Strategy, Operation: Codec, { staged: Staged, db: Shared>, metadata: Option, } impl AnyUnmerkleized where F: Family, E: Context, U: Update, C: Contiguous>, I: UnorderedIndex> + 'static, H: Hasher, S: Strategy, Operation: Codec, { /// Sets the metadata committed by [`merkleize`](UnmerkleizedTrait::merkleize). /// /// The metadata carries over to a batch returned by [`Self::stage`]. pub fn with_metadata(mut self, metadata: U::Value) -> Self { self.metadata = Some(metadata); self } /// Reads a value by key, falling back to applied state. pub async fn get(&self, key: &U::Key) -> Result, Error> { let db = self.db.read().await; self.batch.get(key, &db).await } /// Reads multiple values by key, falling back to applied state. /// /// Returns results in the same order as `keys`. pub async fn get_many(&self, keys: &[&U::Key]) -> Result>, Error> { let db = self.db.read().await; self.batch.get_many(keys, &db).await } /// Reads multiple values and returns a staged batch for the same keys. /// /// Returns results in the same order as `keys`. pub async fn stage( self, keys: &[&U::Key], ) -> Result<(Vec>, AnyStaged), Error> { let Self { batch, db, metadata, } = self; let (values, staged) = { let guard = db.read().await; batch.stage(keys, &guard).await? }; Ok(( values, AnyStaged { staged, db, metadata, }, )) } /// Records an upsert (`Some`) or a delete (`None`) of `key`. pub fn write(mut self, key: U::Key, value: Option) -> Self { self.batch = self.batch.write(key, value); self } } /// A sealed `any` batch with a computed root. pub struct AnyMerkleized where F: Family, E: Context, U: Update, C: Contiguous>, I: UnorderedIndex>, H: Hasher, S: Strategy, Operation: Codec, { inner: Arc>, db: Shared>, } impl Clone for AnyMerkleized where F: Family, E: Context, U: Update, C: Contiguous>, I: UnorderedIndex>, H: Hasher, S: Strategy, Operation: Codec, { fn clone(&self) -> Self { Self { inner: Arc::clone(&self.inner), db: self.db.clone(), } } } impl Deref for AnyUnmerkleized where F: Family, E: Context, U: Update, C: Contiguous>, I: UnorderedIndex>, H: Hasher, S: Strategy, Operation: Codec, { type Target = UnmerkleizedBatch; fn deref(&self) -> &Self::Target { &self.batch } } impl Deref for AnyMerkleized where F: Family, E: Context, U: Update, C: Contiguous>, I: UnorderedIndex>, H: Hasher, S: Strategy, Operation: Codec, { type Target = MerkleizedBatch; fn deref(&self) -> &Self::Target { &self.inner } } impl AnyStaged where F: Family, E: Context, U: Update, C: Contiguous>, I: UnorderedIndex> + 'static, H: Hasher, S: Strategy, Operation: Codec, { /// Sets the metadata committed by [`merkleize`](Self::merkleize), replacing any metadata set /// before staging. pub fn with_metadata(mut self, metadata: U::Value) -> Self { self.metadata = Some(metadata); self } /// The number of staged read slots: the keys passed to `stage` plus every /// [`expand`](Self::expand) range. [`merkleize`](Self::merkleize) takes one [`Slot`] per /// staged read. pub const fn len(&self) -> usize { self.staged.len() } /// Whether no reads are staged. pub const fn is_empty(&self) -> bool { self.staged.is_empty() } /// Expands this staged batch with more reads. /// /// Existing slots keep their positions. Newly read keys are appended to the staged read set /// and assigned the returned range of slots. Expansion does not deduplicate against previously /// staged keys and does not observe values computed for earlier staged slots but not yet /// passed to `merkleize`. pub async fn expand( self, keys: &[&U::Key], ) -> Result<(Range, Vec>, Self), Error> { let Self { staged, db, metadata, } = self; let (range, values, staged) = { let guard = db.read().await; staged.expand(keys, &guard).await? }; Ok(( range, values, Self { staged, db, metadata, }, )) } } impl AnyStaged, S> where F: Family, E: Context, K: Key, V: ValueEncoding, C: Mutable>>, I: UnorderedIndex> + 'static, H: Hasher, S: Strategy, Operation>: Codec, { /// Applies `slots` to staged reads and `unstaged` writes to unread keys, then merkleizes. /// /// `slots` holds one [`Slot`] per staged read, in the order of the keys passed to `stage` /// followed by each [`expand`](Self::expand) range. A key staged in several slots takes its /// last put or delete, and an unstaged write to a staged key wins. Metadata set through /// [`with_metadata`](Self::with_metadata), or before staging, is committed with the batch. /// /// # Panics /// /// Panics if `slots` does not hold one entry per staged read. pub async fn merkleize( self, slots: Vec>, unstaged: Vec<(K, Option)>, ) -> Result, S>, Error> { let Self { staged, db, metadata, } = self; let inner = { let guard = db.read().await; staged .merkleize(slots, unstaged, metadata, &guard, &mut Proportional) .await? }; Ok(AnyMerkleized { inner, db }) } } impl AnyStaged, S> where F: Family, E: Context, K: Key, V: ValueEncoding, C: Mutable>>, I: OrderedIndex> + 'static, H: Hasher, S: Strategy, Operation>: Codec, { /// Applies `slots` to staged reads and `unstaged` writes to unread keys, then merkleizes. /// /// `slots` holds one [`Slot`] per staged read, in the order of the keys passed to `stage` /// followed by each [`expand`](Self::expand) range. A key staged in several slots takes its /// last put or delete, and an unstaged write to a staged key wins. Metadata set through /// [`with_metadata`](Self::with_metadata), or before staging, is committed with the batch. /// /// # Panics /// /// Panics if `slots` does not hold one entry per staged read. pub async fn merkleize( self, slots: Vec>, unstaged: Vec<(K, Option)>, ) -> Result, S>, Error> { let Self { staged, db, metadata, } = self; let inner = { let guard = db.read().await; staged .merkleize(slots, unstaged, metadata, &guard, &mut Proportional) .await? }; Ok(AnyMerkleized { inner, db }) } } impl AnyMerkleized where F: Family, E: Context, U: Update, C: Contiguous>, I: UnorderedIndex> + 'static, H: Hasher, S: Strategy, Operation: Codec, { /// Reads a value by key, falling back to applied state. pub async fn get(&self, key: &U::Key) -> Result, Error> { let db = self.db.read().await; self.inner.get(key, &db).await } /// Reads multiple values by key, falling back to applied state. /// /// Returns results in the same order as `keys`. pub async fn get_many(&self, keys: &[&U::Key]) -> Result>, Error> { let db = self.db.read().await; self.inner.get_many(keys, &db).await } } impl UnmerkleizedTrait for AnyUnmerkleized, S> where F: Family, E: Context, K: Key, V: ValueEncoding, C: Mutable>>, I: UnorderedIndex> + 'static, H: Hasher, S: Strategy, Operation>: Codec, { type Merkleized = AnyMerkleized, S>; type Error = Error; async fn merkleize(self) -> Result> { let db = self.db.read().await; let merkleized = self .batch .merkleize(&db, self.metadata, &mut Proportional) .await?; Ok(AnyMerkleized { inner: merkleized, db: self.db.clone(), }) } } impl UnmerkleizedTrait for AnyUnmerkleized, S> where F: Family, E: Context, K: Key, V: ValueEncoding, C: Mutable>>, I: OrderedIndex> + 'static, H: Hasher, S: Strategy, Operation>: Codec, { type Merkleized = AnyMerkleized, S>; type Error = Error; async fn merkleize(self) -> Result> { let db = self.db.read().await; let merkleized = self .batch .merkleize(&db, self.metadata, &mut Proportional) .await?; Ok(AnyMerkleized { inner: merkleized, db: self.db.clone(), }) } } impl MerkleizedTrait for AnyMerkleized where F: Family, E: Context, U: Update, C: Mutable>, I: UnorderedIndex> + 'static, H: Hasher, S: Strategy, Operation: Codec, AnyUnmerkleized: UnmerkleizedTrait, { type Digest = H::Digest; type Unmerkleized = AnyUnmerkleized; fn root(&self) -> H::Digest { self.inner.root() } fn new_batch(&self) -> Self::Unmerkleized { AnyUnmerkleized { batch: self.inner.new_batch::(), db: self.db.clone(), metadata: None, } } } impl ManagedDb for Db< F, E, FixedJournal>>>, UnorderedIdx>, H, unordered::Update>, ANY_BITMAP_CHUNK_BYTES, S, > where F: Family, E: Context + Spawner, K: Array, V: value::FixedValue, H: Hasher, T: Translator, S: Strategy, { type Unmerkleized = AnyUnmerkleized< F, E, FixedJournal>>>, UnorderedIdx>, H, unordered::Update>, S, >; type Merkleized = AnyMerkleized< F, E, FixedJournal>>>, UnorderedIdx>, H, unordered::Update>, S, >; type Error = Error; type Config = FixedConfig; type SyncTarget = AnySyncTarget; async fn init( context: E, config: Self::Config, expected: Option, ) -> Result, Self::SyncTarget>> { let db = ::init( context, config, expected.as_ref().map(|target| target.range.end()), ) .await .map_err(InitError::Database)?; validate_initialization(db, expected) } fn initial_sync_target() -> Self::SyncTarget { AnySyncTarget::new( initial_root::>, H>(), non_empty_range!(Location::new(0), Location::new(1)), ) } fn new_batch(database: BatchContext<'_, Self>) -> Self::Unmerkleized { let (database, shared) = database.into_parts(); AnyUnmerkleized { batch: database.new_batch(), db: shared, metadata: None, } } fn matches_sync_target(batch: &Self::Merkleized, target: &Self::SyncTarget) -> bool { batch.root() == target.root && *target.range.start() == batch.bounds().inactivity_floor && *target.range.end() == batch.bounds().tip.size } async fn apply(self, batch: Self::Merkleized) -> Result> { self.apply_batch(batch.inner).await } async fn finalize(self) -> Result<(Self, Handle<()>), Error> { self.start_sync().await } async fn prune(self, target: &Self::SyncTarget) -> Result> { self.prune((*target.range.start()).into()).await } fn sync_target(&self) -> Self::SyncTarget { let bounds = self.bounds(); AnySyncTarget::new( self.root(), non_empty_range!(self.sync_boundary(), bounds.end), ) } } impl ManagedDb for Db< F, E, VariableJournal>>>, UnorderedIdx>, H, unordered::Update>, ANY_BITMAP_CHUNK_BYTES, S, > where F: Family, E: Context + Spawner, K: Key, V: value::VariableValue, H: Hasher, T: Translator, S: Strategy, { type Unmerkleized = AnyUnmerkleized< F, E, VariableJournal>>>, UnorderedIdx>, H, unordered::Update>, S, >; type Merkleized = AnyMerkleized< F, E, VariableJournal>>>, UnorderedIdx>, H, unordered::Update>, S, >; type Error = Error; type Config = VariableConfig< T, >> as CodecRead>::Cfg, S, >; type SyncTarget = AnySyncTarget; async fn init( context: E, config: Self::Config, expected: Option, ) -> Result, Self::SyncTarget>> { let db = ::init( context, config, expected.as_ref().map(|target| target.range.end()), ) .await .map_err(InitError::Database)?; validate_initialization(db, expected) } fn initial_sync_target() -> Self::SyncTarget { AnySyncTarget::new( initial_root::>, H>(), non_empty_range!(Location::new(0), Location::new(1)), ) } fn new_batch(database: BatchContext<'_, Self>) -> Self::Unmerkleized { let (database, shared) = database.into_parts(); AnyUnmerkleized { batch: database.new_batch(), db: shared, metadata: None, } } fn matches_sync_target(batch: &Self::Merkleized, target: &Self::SyncTarget) -> bool { batch.root() == target.root && *target.range.start() == batch.bounds().inactivity_floor && *target.range.end() == batch.bounds().tip.size } async fn apply(self, batch: Self::Merkleized) -> Result> { self.apply_batch(batch.inner).await } async fn finalize(self) -> Result<(Self, Handle<()>), Error> { self.start_sync().await } async fn prune(self, target: &Self::SyncTarget) -> Result> { self.prune((*target.range.start()).into()).await } fn sync_target(&self) -> Self::SyncTarget { let bounds = self.bounds(); AnySyncTarget::new( self.root(), non_empty_range!(self.sync_boundary(), bounds.end), ) } } impl StateSyncDb for Db< F, E, FixedJournal>>>, UnorderedIdx>, H, unordered::Update>, ANY_BITMAP_CHUNK_BYTES, S, > where F: Family, E: Context + Spawner, K: Array, V: value::FixedValue, H: Hasher, T: Translator, S: Strategy, R: sync::SourceFor, { type SyncError = sync::Error; async fn sync_db( context: E, config: Self::Config, source: R, target: Self::SyncTarget, tip_updates: ring::Receiver, finish: Option>, reached_target: Option>, sync_config: SyncEngineConfig, ) -> Result { sync_standard_db( context, config, source, target, tip_updates, finish, reached_target, sync_config, ) .await } } impl StateSyncDb for Db< F, E, VariableJournal>>>, UnorderedIdx>, H, unordered::Update>, ANY_BITMAP_CHUNK_BYTES, S, > where F: Family, E: Context + Spawner, K: Key, V: value::VariableValue, H: Hasher, T: Translator, S: Strategy, R: sync::SourceFor, { type SyncError = sync::Error; async fn sync_db( context: E, config: Self::Config, source: R, target: Self::SyncTarget, tip_updates: ring::Receiver, finish: Option>, reached_target: Option>, sync_config: SyncEngineConfig, ) -> Result { sync_standard_db( context, config, source, target, tip_updates, finish, reached_target, sync_config, ) .await } } #[cfg(test)] mod tests { use super::*; use commonware_cryptography::{Sha256, sha256::Digest}; use commonware_parallel::Sequential; use commonware_runtime::{ Runner as _, Supervisor as _, buffer::paged::CacheRef, deterministic, mocks::{DelayedSyncContext, PendingSyncs, drive_pending_syncs, release_pending_syncs}, }; use commonware_storage::{ journal::contiguous::fixed::Config as FixedJournalConfig, merkle::{full::Config as MerkleConfig, mmr}, qmdb::{self, any::unordered::fixed}, translator::TwoCap, }; use commonware_utils::{NZU16, NZU64, NZUsize}; use std::num::{NonZeroU16, NonZeroUsize}; type UnorderedFixedDb = fixed::Db; const PAGE_SIZE: NonZeroU16 = NZU16!(101); const PAGE_CACHE_SIZE: NonZeroUsize = NZUsize!(11); fn fixed_config(suffix: &str) -> FixedConfig { let page_cache = CacheRef::new(PAGE_SIZE, PAGE_CACHE_SIZE); FixedConfig { merkle_config: MerkleConfig { journal_partition: format!("stateful-any-journal-{suffix}"), metadata_partition: format!("stateful-any-metadata-{suffix}"), items_per_blob: NZU64!(11), write_buffer: NZUsize!(1024), replay_buffer: NZUsize!(1024), strategy: Sequential, page_cache: page_cache.clone(), }, journal_config: FixedJournalConfig { partition: format!("stateful-any-log-{suffix}"), items_per_blob: NZU64!(7), page_cache, write_buffer: NZUsize!(1024), replay_buffer: NZUsize!(1024), }, translator: TwoCap, init_cache: Some(NZUsize!(1024)), init_buffer: NZUsize!(1 << 21), init_concurrency: (), } } #[test] fn unmerkleized_batch_refuses_after_competing_finalization() { deterministic::Runner::default().start(|context| async move { let config = fixed_config("unordered-fixed-stale-refusal"); let db = >::init(context.child("db"), config, None) .await .unwrap(); let db = Shared::new("test", db); let key = Sha256::hash(&[b"key"]); let value = Sha256::hash(&[b"winner"]); let pre_finalization = db.new_batch_for_test::<_>().await; let winner = db.new_batch_for_test::<_>().await.write(key, Some(value)); let winner = crate::stateful::db::Unmerkleized::merkleize(winner) .await .unwrap(); let (slot, database) = db.write().await; let database = >::apply(database, winner) .await .unwrap(); let (database, sync) = >::finalize(database) .await .unwrap(); slot.put(database); sync.await.expect("database sync failed"); // `winner` is a sibling built on the same state, so applying it moves the database to // a state outside `pre_finalization`'s chain and reads through `pre_finalization` // return `StaleRead`. assert!(matches!( pre_finalization.get(&key).await, Err(qmdb::Error::StaleRead) )); }); } /// The glue staged wrapper (`AnyUnmerkleized::stage` -> `AnyStaged::expand` -> /// `AnyStaged::merkleize`) must return the same values and root as an explicit `get_many` + /// `write` + `merkleize`, including a staged delete, an unstaged write, and metadata flow (both /// set on the staged handle via `with_metadata` and carried from before staging). This guards /// metadata flow and db-handle pairing through the wrapper. #[test] fn unordered_fixed_staged_merkleize_matches_explicit_writes() { deterministic::Runner::default().start(|context| async move { let config = fixed_config("unordered-fixed-glue-staged"); let db = >::init(context.child("db"), config, None) .await .unwrap(); let db = Shared::new("test", db); let key = |i: u64| Sha256::hash(&[&i.to_be_bytes()]); let val = |i: u64| Sha256::hash(&[&(i + 10_000).to_be_bytes()]); let metadata = Sha256::hash(&[b"metadata"]); // Seed keys 0..50 and persist them. let mut seed = db.new_batch_for_test::<_>().await; for i in 0..50u64 { seed = seed.write(key(i), Some(val(i))); } let merkleized = crate::stateful::db::Unmerkleized::merkleize(seed) .await .unwrap(); { let (slot, database) = db.write().await; let database = >::apply(database, merkleized) .await .unwrap(); let (database, sync) = >::finalize(database) .await .unwrap(); slot.put(database); sync.await.expect("database sync failed"); } // Read set: key(1) updated, key(2) deleted, key(999) missing -> created. let read_keys = [key(1), key(2), key(999)]; let keys: Vec<&Digest> = read_keys.iter().collect(); let slots = vec![Slot::Put(val(1_000)), Slot::Delete, Slot::Put(val(1_001))]; let unstaged = vec![(key(3), Some(val(1_002)))]; // Explicit path. let mut explicit = db.new_batch_for_test::<_>().await; let explicit_values = explicit.get_many(&keys).await.unwrap(); for (key, slot) in read_keys.iter().zip(&slots) { let value = match slot { Slot::Put(value) => Some(*value), Slot::Delete => None, Slot::Keep => unreachable!("every staged read must be written"), }; explicit = explicit.write(*key, value); } for (k, v) in &unstaged { explicit = explicit.write(*k, *v); } let explicit_root = crate::stateful::db::Unmerkleized::merkleize(explicit.with_metadata(metadata)) .await .unwrap() .root(); // Staged path, with metadata set on the staged handle. let staged_batch = db.new_batch_for_test::<_>().await; let split = 2; let (mut staged_values, staged) = staged_batch.stage(&keys[..split]).await.unwrap(); let (range, suffix_values, staged) = staged.expand(&keys[split..]).await.unwrap(); assert_eq!(range, split..keys.len()); assert_eq!(staged.len(), keys.len()); assert!(!staged.is_empty()); staged_values.extend(suffix_values); let staged_root = staged .with_metadata(metadata) .merkleize(slots.clone(), unstaged.clone()) .await .unwrap() .root(); assert_eq!(explicit_values, staged_values); assert_eq!(explicit_root, staged_root); // Metadata set before staging must be carried through to staged merkleize. let carried_batch = db.new_batch_for_test::<_>().await.with_metadata(metadata); let (carried_values, staged) = carried_batch.stage(&keys).await.unwrap(); let carried_root = staged .merkleize(slots.clone(), unstaged.clone()) .await .unwrap() .root(); assert_eq!(explicit_values, carried_values); assert_eq!(explicit_root, carried_root); }); } type DelayedFixedDb = fixed::Db< mmr::Family, DelayedSyncContext, Digest, Digest, Sha256, TwoCap, Sequential, >; #[test] fn apply_does_not_finalize() { deterministic::Runner::default().start(|context| async move { let pending = PendingSyncs::default(); let delayed = DelayedSyncContext { inner: context, pending: pending.clone(), }; let config = fixed_config("unordered-fixed-deferred"); let db = drive_pending_syncs( &pending, >::init(delayed.child("db"), config, None), ) .await .unwrap(); let db = Shared::new("test", db); let key = Sha256::hash(&[b"key"]); let value = Sha256::hash(&[b"value"]); let batch = db.new_batch_for_test::<_>().await.write(key, Some(value)); let merkleized = crate::stateful::db::Unmerkleized::merkleize(batch) .await .unwrap(); let starts = pending.starts(); let (slot, database) = db.write().await; let database = >::apply(database, merkleized) .await .unwrap(); slot.put(database); assert_eq!(pending.starts(), starts, "apply must not start durability"); { let guard = db.read().await; assert_eq!(guard.get(&key).await.unwrap(), Some(value)); } let (slot, database) = db.write().await; let (database, sync) = >::finalize(database) .await .unwrap(); slot.put(database); assert!( pending.starts() > pending.completions(), "finalize must leave its flush parked", ); release_pending_syncs(&pending); drive_pending_syncs(&pending, sync) .await .expect("flush must succeed once released"); }); } }