Skip to main content

reddb_server/
physical.rs

1//! Physical storage design primitives for RedDB's deterministic on-disk layout.
2
3use std::collections::BTreeMap;
4use std::fs;
5use std::io;
6use std::path::{Path, PathBuf};
7use std::time::{SystemTime, UNIX_EPOCH};
8
9use crate::api::{CatalogSnapshot, CollectionStats, RedDBOptions, SchemaManifest, StorageMode};
10use crate::index::IndexKind;
11use crate::serde_json::{Map, Value as JsonValue};
12
13pub const DEFAULT_GRID_BLOCK_SIZE: usize = 512 * 1024;
14pub const DEFAULT_PAGE_SIZE: usize = 4096;
15pub use reddb_file::layout::PHYSICAL_METADATA_BINARY_EXTENSION;
16pub use reddb_file::{
17    fold_dwb_into_wal_enabled, meta_json_sidecar_enabled, seqn_journal_enabled,
18    seqn_journal_retention, set_fold_dwb_into_wal_enabled, set_meta_json_sidecar_enabled,
19    set_seqn_journal_enabled, set_seqn_journal_retention, BlockReference, ExportDescriptor,
20    ManifestEvent, ManifestEventKind, ManifestPointers, PhysicalAnalyticsJob,
21    PhysicalGraphProjection, PhysicalTreeDefinition, SnapshotDescriptor, SuperblockHeader,
22    DEFAULT_METADATA_JOURNAL_RETENTION, DEFAULT_SUPERBLOCK_COPIES,
23    OPT_IN_METADATA_JOURNAL_RETENTION, PHYSICAL_METADATA_PROTOCOL_VERSION,
24};
25pub const DEFAULT_MANIFEST_EVENT_HISTORY: usize = 256;
26
27#[derive(Debug, Clone, Copy, PartialEq, Eq)]
28pub enum PhysicalMetadataSource {
29    Binary,
30    BinaryJournal,
31    Json,
32}
33
34impl PhysicalMetadataSource {
35    pub fn as_str(self) -> &'static str {
36        match self {
37            Self::Binary => "binary",
38            Self::BinaryJournal => "binary_journal",
39            Self::Json => "json",
40        }
41    }
42}
43#[derive(Debug, Clone, Copy, PartialEq, Eq)]
44pub enum CompactionPolicy {
45    Incremental,
46    Manual,
47}
48
49#[derive(Debug, Clone)]
50pub struct WalPolicy {
51    pub auto_checkpoint_pages: u32,
52    pub fsync_on_commit: bool,
53    pub ring_buffer_bytes: u64,
54}
55
56impl Default for WalPolicy {
57    fn default() -> Self {
58        Self {
59            auto_checkpoint_pages: 1000,
60            fsync_on_commit: true,
61            ring_buffer_bytes: 64 * 1024 * 1024,
62        }
63    }
64}
65
66#[derive(Debug, Clone)]
67pub struct GridLayout {
68    pub block_size: usize,
69    pub page_size: usize,
70    pub superblock_copies: u8,
71}
72
73impl Default for GridLayout {
74    fn default() -> Self {
75        Self {
76            block_size: DEFAULT_GRID_BLOCK_SIZE,
77            page_size: DEFAULT_PAGE_SIZE,
78            superblock_copies: DEFAULT_SUPERBLOCK_COPIES,
79        }
80    }
81}
82
83#[derive(Debug, Clone)]
84pub struct PhysicalLayout {
85    pub mode: StorageMode,
86    pub grid: GridLayout,
87    pub wal: WalPolicy,
88    pub compaction: CompactionPolicy,
89}
90
91impl PhysicalLayout {
92    pub fn from_options(options: &RedDBOptions) -> Self {
93        Self {
94            mode: options.mode,
95            grid: GridLayout::default(),
96            wal: WalPolicy {
97                auto_checkpoint_pages: options.auto_checkpoint_pages,
98                ..WalPolicy::default()
99            },
100            compaction: CompactionPolicy::Incremental,
101        }
102    }
103
104    pub fn is_persistent(&self) -> bool {
105        self.mode == StorageMode::Persistent
106    }
107}
108
109pub(crate) use reddb_engine::catalog::{
110    is_collection_contract_decode_error, CollectionContractDecodeError,
111};
112pub use reddb_engine::catalog::{CollectionContract, ContractOrigin, DeclaredColumnContract};
113
114/// Canonical artifact lifecycle states.
115///
116/// State machine transitions:
117/// ```text
118///   Declared ──► Building ──► Ready ──► Stale ──► RequiresRebuild
119///       │            │          │                       │
120///       │            ▼          ▼                       │
121///       │         Failed    Disabled                    │
122///       │            │                                  │
123///       └────────────┴──────────────────────────────────┘
124///                    (rebuild restarts from Building)
125/// ```
126#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
127pub enum ArtifactState {
128    /// Index declared but never materialized.
129    Declared,
130    /// Artifact is being built or rebuilt.
131    Building,
132    /// Artifact is materialized and queryable.
133    Ready,
134    /// Artifact is explicitly disabled by the operator.
135    Disabled,
136    /// Underlying data changed; artifact is out of date.
137    Stale,
138    /// Build or warmup failed; manual intervention may be needed.
139    Failed,
140    /// Artifact must be rebuilt before it can serve reads.
141    RequiresRebuild,
142}
143
144impl ArtifactState {
145    /// Parse from the legacy string representation stored in physical metadata.
146    pub fn from_build_state(s: &str, enabled: bool) -> Self {
147        if !enabled {
148            return Self::Disabled;
149        }
150        match s {
151            "ready" => Self::Ready,
152            "building" | "catalog-derived" | "metadata-only" | "artifact-published"
153            | "registry-loaded" => Self::Building,
154            "stale" => Self::Stale,
155            "failed" => Self::Failed,
156            "requires_rebuild" | "requires-rebuild" => Self::RequiresRebuild,
157            _ => Self::Declared,
158        }
159    }
160
161    /// Canonical string representation for storage and API surfaces.
162    pub fn as_str(&self) -> &'static str {
163        match self {
164            Self::Declared => "declared",
165            Self::Building => "building",
166            Self::Ready => "ready",
167            Self::Disabled => "disabled",
168            Self::Stale => "stale",
169            Self::Failed => "failed",
170            Self::RequiresRebuild => "requires_rebuild",
171        }
172    }
173
174    /// Whether this artifact is safe for query reads.
175    pub fn is_queryable(&self) -> bool {
176        matches!(self, Self::Ready)
177    }
178
179    /// Whether a rebuild operation is valid from this state.
180    pub fn can_rebuild(&self) -> bool {
181        matches!(
182            self,
183            Self::Declared | Self::Stale | Self::Failed | Self::RequiresRebuild
184        )
185    }
186
187    /// Whether this state indicates the artifact needs attention.
188    pub fn needs_attention(&self) -> bool {
189        matches!(self, Self::Failed | Self::RequiresRebuild | Self::Stale)
190    }
191}
192
193impl std::fmt::Display for ArtifactState {
194    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
195        f.write_str(self.as_str())
196    }
197}
198
199#[derive(Debug, Clone)]
200pub struct PhysicalIndexState {
201    pub name: String,
202    pub kind: IndexKind,
203    pub collection: Option<String>,
204    pub enabled: bool,
205    pub entries: usize,
206    pub estimated_memory_bytes: u64,
207    pub last_refresh_ms: Option<u128>,
208    pub backend: String,
209    pub artifact_kind: Option<String>,
210    pub artifact_root_page: Option<u32>,
211    pub artifact_checksum: Option<u64>,
212    pub build_state: String,
213}
214
215impl PhysicalIndexState {
216    /// Canonical artifact lifecycle state derived from physical state.
217    pub fn artifact_state(&self) -> ArtifactState {
218        ArtifactState::from_build_state(&self.build_state, self.enabled)
219    }
220}
221
222/// A single persisted hypertable chunk. Mirror of
223/// `storage::timeseries::ChunkMeta`, flattened for the metadata
224/// sidecar so the registry's routing spine survives a restart
225/// (issue #866). `start_ns` plus the owning hypertable name is the
226/// chunk's stable identity.
227#[derive(Debug, Clone)]
228pub struct PhysicalHypertableChunk {
229    pub start_ns: u64,
230    pub end_ns_exclusive: u64,
231    pub row_count: u64,
232    pub min_ts_ns: u64,
233    pub max_ts_ns: u64,
234    pub sealed: bool,
235    pub ttl_override_ns: Option<u64>,
236    /// Columnar-vs-row migration discriminant — mirror of
237    /// `ChunkMeta.columnar_page` (PRD #850, Phase 1). `Some` → the chunk's
238    /// `RDCC` `ColumnBlock` location; `None` → legacy row-stored. Absent on
239    /// sidecars written before the feature, decoding to `None`.
240    pub columnar_page: Option<crate::storage::engine::PageLocation>,
241    /// Columnar blocks are derived, rebuildable artifacts (ADR 0069), not
242    /// source-of-truth row data. Backup/restore may skip or rebuild them.
243    pub columnar_derived: bool,
244}
245
246/// A persisted hypertable spec plus all of its chunks. Stored in the
247/// physical metadata sidecar alongside collection contracts so chunk
248/// bounds / routing / TTL are recovered identically after a restart
249/// — the same durability path the rest of the catalog already uses,
250/// not a parallel one (issue #866).
251#[derive(Debug, Clone)]
252pub struct PhysicalHypertable {
253    pub name: String,
254    pub time_column: String,
255    pub chunk_interval_ns: u64,
256    pub default_ttl_ns: Option<u64>,
257    pub chunks: Vec<PhysicalHypertableChunk>,
258}
259
260#[derive(Debug, Clone)]
261pub struct PhysicalMetadataFile {
262    pub protocol_version: String,
263    pub generated_at_unix_ms: u128,
264    pub last_loaded_from: Option<String>,
265    pub last_healed_at_unix_ms: Option<u128>,
266    pub manifest: SchemaManifest,
267    pub catalog: CatalogSnapshot,
268    pub manifest_events: Vec<ManifestEvent>,
269    pub indexes: Vec<PhysicalIndexState>,
270    pub graph_projections: Vec<PhysicalGraphProjection>,
271    pub analytics_jobs: Vec<PhysicalAnalyticsJob>,
272    pub tree_definitions: Vec<PhysicalTreeDefinition>,
273    pub collection_ttl_defaults_ms: BTreeMap<String, u64>,
274    pub collection_contracts: Vec<CollectionContract>,
275    /// Persisted hypertable chunk spine (issue #866). Empty on legacy
276    /// sidecars written before the feature and for non-hypertable
277    /// databases.
278    pub hypertables: Vec<PhysicalHypertable>,
279    pub exports: Vec<ExportDescriptor>,
280    pub superblock: SuperblockHeader,
281    pub snapshots: Vec<SnapshotDescriptor>,
282}
283
284mod helpers;
285mod json_codec;
286mod metadata_file;
287pub mod shm;
288
289pub(crate) use reddb_engine::catalog::codec::{
290    deserialize_collection_contracts, serialize_collection_contracts,
291};
292
293pub use self::shm::{
294    provision_shm, read_shm_header, set_shm_provisioning_enabled, shm_path_for,
295    shm_provisioning_enabled, ShmHandle, ShmHeader, ShmProvisionState, SHM_FILE_SIZE,
296    SHM_HEADER_SIZE, SHM_MAGIC, SHM_VERSION,
297};
298
299use self::helpers::*;
300use self::json_codec::*;