artifactize_core/
store.rs

1//! The state database and storage maintenance.
2
3use std::path::{Path, PathBuf};
4
5use crate::{platform, workspace::canonical_target};
6
7mod blocked;
8pub mod definitions;
9mod executions;
10pub(crate) mod history;
11mod human;
12pub use human::HumanClaim;
13mod catalog;
14pub(crate) mod connection;
15pub mod prune;
16pub(crate) mod receipts;
17mod result;
18mod run_record;
19mod runs;
20mod snapshots;
21mod touched;
22mod unreadable;
23mod validation;
24pub use unreadable::{Unreadable, read_unreadable};
25mod lifecycle;
26pub(crate) mod migrations;
27pub(crate) mod run_control;
28pub use lifecycle::{ExecutionState, GreenResult, PendingRequest, RequestState, RunState};
29pub use result::{ExecutionResult, Field as ResultField, RuntimeResult};
30pub use snapshots::{
31    ArtifactValidation, Blocker, ChildIdentity, EvalValidation, HumanDefinition, Validation,
32    ValidationSnapshot,
33};
34mod wait_timeout;
35pub use catalog::{CatalogRun, Signoff, read_catalog};
36pub use runs::{RunSummary, read_runs, read_scoped_runs};
37pub(crate) mod reader;
38mod requests;
39pub(crate) use executions::read_profiled_executions;
40pub use executions::{
41    Capacity, Claim, Execution, ExecutionOptions, Origin, Producer, Provenance,
42    read_keyed_executions, read_latest_cached,
43};
44#[cfg(any(test, feature = "test-support"))]
45pub use receipts::OwnedReceipts;
46pub(crate) use receipts::regular_files as check_probe_files;
47pub use receipts::{
48    DATABASE, EARLIER_STATE, LastRequest, QueueReason, Receipts, Request, Run, RunView,
49    StoppedBackend, read_latest_requests, read_run, read_state_id, schema_error, state_schema,
50};
51pub use requests::read_request_repo;
52pub use requests::{
53    RequestView, read_original, read_request, read_request_batch, read_requests, read_requests_in,
54    read_run_requests, read_session_request, read_waiting, read_waiting_originals,
55};
56
57/// How long one database operation waits for the state other processes are writing before it
58/// fails. Several Runs, the monitor and MCP servers share one state; a Run's write waits behind
59/// theirs rather than failing the Run, while a lock that never frees still ends in an error.
60pub(crate) const SQLITE_BUSY_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(30);
61
62/// How often a busy operation tries again. SQLite's own busy timeout backs off to 100 ms
63/// between tries, so under steady writes from other connections, which try every 1–2 ms when
64/// they start waiting, a long waiter rarely finds the lock free and runs out its time while
65/// the others commit (#372). Trying at this pace keeps every waiter in the race.
66const SQLITE_BUSY_POLL: std::time::Duration = std::time::Duration::from_millis(1);
67
68/// Make `db` wait for a busy database, trying every [`SQLITE_BUSY_POLL`] for up to
69/// [`SQLITE_BUSY_TIMEOUT`].
70pub(crate) fn wait_when_busy(db: &rusqlite::Connection) -> rusqlite::Result<()> {
71    db.busy_handler(Some(busy))
72}
73
74/// SQLite's busy handler: `tries` counts the earlier calls for this one wait, which runs on
75/// the calling thread, so the wait's start is kept per thread.
76fn busy(tries: i32) -> bool {
77    thread_local! {
78        static SINCE: std::cell::Cell<std::time::Instant> = std::cell::Cell::new(std::time::Instant::now());
79    }
80    let now = std::time::Instant::now();
81    if tries == 0 {
82        SINCE.set(now);
83    }
84    if now.duration_since(SINCE.get()) >= SQLITE_BUSY_TIMEOUT {
85        return false;
86    }
87    std::thread::sleep(SQLITE_BUSY_POLL);
88    true
89}
90
91/// Allow four short backoffs for a pending-delete handle to close, bounding the extra
92/// latency of a real file-access failure to 80ms. This is a retry policy, not a test deadline.
93const TRANSIENT_ACCESS_RETRIES: u32 = 5;
94const TRANSIENT_ACCESS_BACKOFF: std::time::Duration = std::time::Duration::from_millis(20);
95
96/// Retry `attempt` while `transient` says its error can clear on its own, before giving up
97/// with its last error. `transient` decides alone: a non-transient error, or the bound run out,
98/// both stop the loop and return that error immediately, never retried further.
99fn retry<T, E>(
100    mut attempt: impl FnMut() -> Result<T, E>,
101    transient: impl Fn(&E) -> bool,
102    mut before_retry: impl FnMut(),
103) -> Result<T, E> {
104    for remaining in (0..TRANSIENT_ACCESS_RETRIES).rev() {
105        match attempt() {
106            Ok(value) => return Ok(value),
107            Err(error) if remaining > 0 && transient(&error) => {
108                before_retry();
109            }
110            Err(error) => return Err(error),
111        }
112    }
113    unreachable!("the loop above always returns by its last attempt")
114}
115
116/// Retry `attempt` past [`platform::transient_file_access`] (Windows transiently denying a
117/// metadata or open probe while another handle finishes closing a just-deleted WAL sidecar)
118/// before giving up with its last error. Every read-only state open reaches this through
119/// [`receipts::check_files`]/[`check_probe_files`], which every reader calls first, so this is
120/// the one place the retry needs to live.
121pub(crate) fn retry_transient_access<T>(
122    attempt: impl FnMut() -> std::io::Result<T>,
123    before_retry: impl FnMut(),
124) -> std::io::Result<T> {
125    retry(attempt, platform::transient_file_access, before_retry)
126}
127
128#[cfg(test)]
129mod retry_tests {
130    use std::{cell::Cell, io};
131
132    #[test]
133    fn file_probe_returns_success_and_permanent_errors_without_backoff() {
134        let calls = Cell::new(0);
135        let result = super::retry_transient_access(
136            || {
137                calls.set(calls.get() + 1);
138                Ok(7)
139            },
140            || panic!("successful probes need no backoff"),
141        );
142        assert_eq!(result.unwrap(), 7);
143        assert_eq!(calls.get(), 1);
144
145        let calls = Cell::new(0);
146        let result = super::retry_transient_access(
147            || {
148                calls.set(calls.get() + 1);
149                Err::<(), _>(io::Error::other("permanent probe failure"))
150            },
151            || panic!("permanent errors cannot be retried"),
152        );
153        assert_eq!(result.unwrap_err().to_string(), "permanent probe failure");
154        assert_eq!(calls.get(), 1);
155    }
156}
157
158/// Resolve the single state directory without creating it.
159pub fn state_dir(explicit: Option<&Path>) -> Result<PathBuf, String> {
160    let path = match explicit {
161        Some(path) => path.to_path_buf(),
162        None => state_home().map_err(|e| e.to_string())?,
163    };
164    canonical_target(&path).map_err(|e| e.to_string())
165}
166
167/// SQLite `user_version`; schema 9 keeps a Run's definitions in their own row, written once,
168/// and numbers every change of a Run, request or execution row; schema 10 adds the open
169/// WAITING_HUMAN Run and names a Run that ended with only sign-offs left YELLOW. Versions from
170/// schema 6 upgrade additively.
171pub const STATE_SCHEMA_VERSION: u32 = 10;
172
173#[derive(Debug, thiserror::Error)]
174#[error("cannot resolve state home: set {}", state_home_variables())]
175pub struct StateHomeError;
176
177/// The variables `state_home` reads, in order, for the error that names them.
178fn state_home_variables() -> String {
179    let names: Vec<&str> = std::iter::once("ARTIFACTIZE_STATE_HOME")
180        .chain(platform::STATE_VARIABLES.iter().copied())
181        .chain([platform::HOME_VARIABLE])
182        .collect();
183    let (last, rest) = names.split_last().expect("constant names");
184    format!("{}, or {last}", rest.join(", "))
185}
186
187/// Resolve the default state home without creating directories; empty variables are ignored.
188/// The platform's state variables come after `ARTIFACTIZE_STATE_HOME` and before the home.
189pub fn state_home() -> Result<PathBuf, StateHomeError> {
190    platform::state_directory().ok_or(StateHomeError)
191}
192
193#[cfg(test)]
194#[path = "store/file_tests.rs"]
195mod file_tests;
artifactize_core/store.rs

artifactize_core/
store.rs
