1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
//! The per-shard tiering state (tier backend builds only).
use kevy_vlog::Vlog;
use crate::{EvictionPolicy, SmallBytes};
/// Per-shard tiering state — present only when tiering is enabled
/// (`tier: Option<TierState>`; `None` = today's paths, the A1 gate's
/// precondition).
#[derive(Debug)]
pub(crate) struct TierState {
pub(crate) vlog: Vlog,
pub(crate) budget: u64,
/// Demotion victim scoring (RFC §7: tiered-lru default).
pub(crate) policy: EvictionPolicy,
pub(crate) demotions_total: u64,
/// Demote-sampler backoff: ticks left to skip before
/// the next over-target sample walk. "Idempotent is not
/// convergent" — a store that is over target with nothing left
/// to spill (every spillable value already cold, or the floor
/// alone exceeds the budget so `effective_target == 0`) used to
/// re-walk the sample window every tick forever.
pub(crate) tick_wait: u32,
/// Current backoff width: doubles on every dry tick batch up
/// to [`crate::tier_demote::BACKOFF_CEILING_TICKS`], resets to
/// 0 on any demotion (tick or write path — the write path
/// always samples immediately, so a fresh spillable value
/// never waits out the window).
pub(crate) tick_skip: u32,
/// Where the demotion sampler starts its next window. It moves
/// past each window it walks, so the sampler sweeps the table
/// with or without accesses: a start drawn from the access clock
/// stood still while nothing but a backfill ran, and demotion
/// stalled once that one window had gone cold.
pub(crate) hand: usize,
pub(crate) promotions_total: u64,
/// Every vlog record read (serve, promote, peek) — the
/// WRONGTYPE-without-read proof counter.
pub(crate) preads_total: u64,
/// Record reads made by NO-PROMOTE peeks only: hydration,
/// backfill, digest, scope-move. One per cold ROW — the
/// preads==rows (not rows×fields) proof counter.
pub(crate) peek_preads_total: u64,
/// Batched cold-read submissions: one per
/// [`Store::peek_hash_rows`] page with ≥1 cold row, weighted by
/// the reader's kernel submission count — the one-batch-per-page
/// proof counter.
pub(crate) batch_submissions_total: u64,
pub(crate) cold_keys: u64,
pub(crate) cold_bytes: u64,
/// Largest value weight demotion may spill (bytes; 0 =
/// unlimited). Bounds the pread-under-shard-lock hold time on
/// the embedded RwLock shape (RFC §7: embedded default 256 KiB,
/// server unlimited) — an over-cap value simply stays hot.
pub(crate) max_spill: u64,
/// Index/view memory floor (Σ segment `approx_bytes` on this
/// shard), fed per shard tick by [`Store::set_tier_reserved`].
/// Subtracted from the demote watermark: the
/// premium fixed layer demotion can never reclaim.
pub(crate) reserved_bytes: u64,
/// This shard's share of what the process holds live outside
/// `used_memory` and the index floor (buffers, rings, allocator
/// overhead), measured by the serving layer. Lowers the target
/// like the index floor does.
pub(crate) overhead_bytes: u64,
/// What the keyspace table's next growth will add, set aside while
/// the table is within an eighth of it.
pub(crate) growth_reserve: u64,
/// RAM the cold stubs themselves cost (Σ per cold key of
/// `ENTRY_OVERHEAD + key heap bytes`), maintained incrementally
/// at demote / promote / DEL-of-cold / RENAME / FLUSHALL. A gauge:
/// the stubs are already inside `used_memory`, so the demote
/// target does not subtract it.
pub(crate) stub_bytes: u64,
/// Cold stubs RENAMEd away from their record's embedded key:
/// `(file_id, offset) → current key`. Rename moves the stub
/// without a pread, so the on-disk key goes stale; compaction's
/// `is_live`/`moved` consult this map on a primary-key miss.
/// Usually empty; entries die with their stub.
pub(crate) renames: std::collections::HashMap<(u32, u64), SmallBytes>,
}