timelake 0.7.0

The open time-series lakehouse: fast queries on your own object storage, with no database and no maintenance jobs
Documentation 
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
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315

//! Materialized rollups end to end (SPEC §11.1): created from SQL, refreshed, and routed to,
//! always giving the answer the raw table gives.

#![allow(clippy::unwrap_used, reason = "test code")]

mod common;

use arrow::array::RecordBatch;
use common::{DAY, T0, batch, opts, rows, scalar_i64, temp_lake};
use timelake::{AppendOptions, Lake, Session, property};

const HOUR: i64 = DAY / 24;
const MINUTE: i64 = HOUR / 60;

const CREATE: &str = "CREATE MATERIALIZED ROLLUP iot.r_1h AS
    SELECT device, time_bucket(INTERVAL '1 hour', time) AS bucket,
           avg(value) AS avg_value, min(value), max(value), sum(value), count(*) AS n, count(value),
           first_value(value ORDER BY time) AS first, last_value(value ORDER BY time) AS last,
           m4(time, value) AS points, time_weighted_mean(time, value) AS twa,
           time_weighted_mean(time, value, 'linear') AS twa_linear
    FROM iot.r GROUP BY device, bucket
    WITH (refresh_interval = 'off')";

/// Queries the rollup can answer (all or part of), and one it can't.
const QUERIES: &[(&str, bool)] = &[
    ("SELECT device, time_bucket(INTERVAL '1 hour', time) AS h, avg(value), count(*) FROM iot.r GROUP BY 1, 2", true),
    ("SELECT device, time_bucket(INTERVAL '1 hour', time) AS h, m4(time, value) FROM iot.r GROUP BY 1, 2", true),
    (
        "SELECT device, time_bucket(INTERVAL '1 hour', time) AS h, time_weighted_mean(time, value),
                time_weighted_mean(time, value, 'linear') FROM iot.r GROUP BY 1, 2",
        true,
    ),
    (
        "SELECT time_bucket(INTERVAL '1 day', time) AS d, device, time_weighted_mean(time, value, 'linear')
         FROM iot.r WHERE time >= TIMESTAMP '2026-01-01 05:30:00' GROUP BY 1, 2",
        true,
    ),
    (
        "SELECT time_bucket(INTERVAL '1 hour', time) AS h, device, max(value) AS m, sum(value) FROM iot.r
         WHERE time >= TIMESTAMP '2026-01-01 05:30:00' AND time < TIMESTAMP '2026-01-02 03:00:00' GROUP BY 1, 2",
        true,
    ),
    (
        "SELECT device, date_bin(INTERVAL '60 minutes', time) AS h, min(value), count(value),
                first_value(value ORDER BY time), last_value(value ORDER BY time)
         FROM iot.r WHERE device IN (1, 2) AND time > TIMESTAMP '2026-01-01 10:00:00' GROUP BY 1, 2",
        true,
    ),
    (
        "SELECT device, time_bucket(INTERVAL '1 hour', time) AS h, avg(value) FROM iot.r
         WHERE time > now() - INTERVAL '400 days' GROUP BY 1, 2",
        true,
    ),
    (
        "SELECT t.device, time_bucket(INTERVAL '1 hour', t.time) AS h, max(t.value) FROM iot.r t
         GROUP BY 1, 2 HAVING max(t.value) > 15000",
        true,
    ),
    // Wider buckets and fewer keys: the rollup's hours merged.
    (
        "SELECT device, time_bucket(INTERVAL '1 day', time) AS d, avg(value), min(value), max(value), count(*),
                sum(value), count(value), first_value(value ORDER BY time), last_value(value ORDER BY time),
                m4(time, value)
         FROM iot.r GROUP BY 1, 2",
        true,
    ),
    (
        "SELECT time_bucket(INTERVAL '6 hours', time) AS b, count(*) AS n, avg(value), max(value) FROM iot.r
         GROUP BY 1",
        true,
    ),
    (
        "SELECT device, time_bucket(INTERVAL '3 hours', time) AS b, avg(value) FROM iot.r
         WHERE time >= TIMESTAMP '2026-01-01 04:30:00' AND time < TIMESTAMP '2026-01-02 20:00:00' GROUP BY 1, 2",
        true,
    ),
    ("SELECT device, time_bucket(INTERVAL '90 minutes', time) AS h, avg(value) FROM iot.r GROUP BY 1, 2", false),
    ("SELECT device, time_bucket(INTERVAL '30 minutes', time) AS h, avg(value) FROM iot.r GROUP BY 1, 2", false),
    (
        "SELECT device, time_bucket(INTERVAL '1 hour', time) AS h, avg(value) FROM iot.r WHERE value > 5 GROUP BY 1, 2",
        false,
    ),
];

/// A query's rows, sorted, as text.
async fn answer(s: &Session, sql: &str) -> String {
    let batches: Vec<RecordBatch> = rows(s, &format!("SELECT * FROM ({sql}) ORDER BY 1, 2")).await;
    arrow::util::pretty::pretty_format_batches(&batches).unwrap().to_string()
}

async fn reads_rollup(s: &Session, sql: &str) -> bool {
    let plan = s.sql(sql).await.unwrap().into_optimized_plan().unwrap();
    plan.display_indent().to_string().contains("r_1h")
}

/// Every query gives the same rows with routing on and off, and is routed when it should be.
async fn check_answers(lake: &Lake) {
    let routed = Session::new(lake.clone());
    let unrouted = Session::new(lake.clone());
    unrouted.sql("SET timelake.rollup_routing = false").await.unwrap();
    for &(sql, routable) in QUERIES {
        assert_eq!(reads_rollup(&routed, sql).await, routable, "{sql}");
        assert!(!reads_rollup(&unrouted, sql).await, "{sql}");
        assert_eq!(answer(&routed, sql).await, answer(&unrouted, sql).await, "{sql}");
    }
}

async fn lake_with_rollup() -> (tempfile::TempDir, Lake) {
    let (dir, lake) = temp_lake();
    let t = lake.create_table("iot.r", &common::schema(), opts()).await.unwrap();
    // Two days, a reading a minute from 4 devices; whole numbers, so sums are exact.
    let b = batch(T0, 2 * 24 * 60, MINUTE, 4, |d, i| f64::from(d) * 10_000.0 + i as f64);
    t.append(&[b], AppendOptions::default()).await.unwrap();
    let s = Session::new(lake.clone());
    let created = rows(&s, CREATE).await;
    let text = arrow::util::pretty::pretty_format_batches(&created).unwrap().to_string();
    assert!(text.contains("iot.r_1h") && text.contains("192"), "4 devices × 48 hours\n{text}");
    (dir, lake)
}

#[tokio::test(flavor = "multi_thread")]
async fn rollups_answer_like_the_raw_table() {
    let (_dir, lake) = lake_with_rollup().await;
    let raw = lake.table("iot.r").await.unwrap();
    assert_eq!(raw.contract().rollups, ["iot.r_1h"]);
    let s = Session::new(lake.clone());
    assert_eq!(scalar_i64(&s, "SELECT count(*) FROM iot.r_1h").await, 192);
    // Columns in the order the query named them.
    let rollup = lake.table("iot.r_1h").await.unwrap();
    let schema = rollup.schema();
    let names: Vec<&str> = schema.fields().iter().map(|f| f.name().as_str()).collect();
    assert_eq!(
        names,
        [
            "device",
            "bucket",
            "avg_value",
            "min_value",
            "max_value",
            "sum_value",
            "n",
            "count_value",
            "first",
            "last",
            "points",
            "twa",
            "twa_linear"
        ]
    );
    check_answers(&lake).await;
}

/// A rollup with averages and counts but no sums answers wider buckets by weighting each
/// bucket's average by its count.
#[tokio::test(flavor = "multi_thread")]
async fn averages_merge_from_average_and_count() {
    let (_dir, lake) = temp_lake();
    let t = lake.create_table("iot.r", &common::schema(), opts()).await.unwrap();
    let b = batch(T0, 2 * 24 * 60, MINUTE, 4, |d, i| f64::from(d) * 10_000.0 + i as f64);
    t.append(&[b], AppendOptions::default()).await.unwrap();
    let s = Session::new(lake.clone());
    rows(
        &s,
        "CREATE MATERIALIZED ROLLUP iot.r_1h AS SELECT device, time_bucket(INTERVAL '1 hour', time) AS bucket,
         avg(value) AS mean, count(value) AS n FROM iot.r GROUP BY device, bucket WITH (refresh_interval = 'off')",
    )
    .await;
    let sql = "SELECT device, time_bucket(INTERVAL '1 day', time) AS d, avg(value) FROM iot.r GROUP BY 1, 2";
    let unrouted = Session::new(lake.clone());
    unrouted.sql("SET timelake.rollup_routing = false").await.unwrap();
    assert!(reads_rollup(&s, sql).await);
    assert_eq!(answer(&s, sql).await, answer(&unrouted, sql).await);
}

/// A daily rollup next to an hourly one is computed from the hourly rollup where that is up to
/// date, and holds what computing it from the raw rows gives: at creation and after a refresh.
#[tokio::test(flavor = "multi_thread")]
async fn a_wider_rollup_is_computed_from_a_narrower_one() {
    let (_dir, lake) = lake_with_rollup().await;
    let s = Session::new(lake.clone());
    rows(
        &s,
        "CREATE MATERIALIZED ROLLUP iot.r_1d AS SELECT device, time_bucket(INTERVAL '1 day', time) AS bucket,
         avg(value) AS avg_value, max(value), count(*) AS n FROM iot.r GROUP BY device, bucket
         WITH (refresh_interval = 'off')",
    )
    .await;
    let expected = "SELECT device, time_bucket(INTERVAL '1 day', time) AS bucket, avg(value) AS avg_value,
                           max(value) AS max_value, count(*) AS n
                    FROM iot.r GROUP BY 1, 2";
    let stored = "SELECT device, bucket, avg_value, max_value, n FROM iot.r_1d";
    let unrouted = Session::new(lake.clone());
    unrouted.sql("SET timelake.rollup_routing = false").await.unwrap();
    assert_eq!(answer(&unrouted, stored).await, answer(&unrouted, expected).await);

    // A third day, then a refresh of both: the hourly rollup first, the daily from it.
    let t = lake.table("iot.r").await.unwrap();
    let b = batch(T0 + 2 * DAY, 24 * 60, MINUTE, 4, |d, i| f64::from(d) * 10_000.0 + i as f64);
    t.append(&[b], AppendOptions { maintain: Some(false), ..AppendOptions::default() }).await.unwrap();
    rows(&s, "REFRESH MATERIALIZED ROLLUP iot.r_1h").await;
    rows(&s, "REFRESH MATERIALIZED ROLLUP iot.r_1d").await;
    assert_eq!(scalar_i64(&unrouted, "SELECT count(*) FROM iot.r_1d").await, 12);
    assert_eq!(answer(&unrouted, stored).await, answer(&unrouted, expected).await);
}

#[tokio::test(flavor = "multi_thread")]
async fn changes_since_the_refresh_come_from_the_raw_table() {
    let (_dir, lake) = lake_with_rollup().await;
    let t = lake.table("iot.r").await.unwrap();
    // A newer write of an old reading (dedup keeps it), and a new hour, without refreshing.
    let late = batch(T0 + 3 * HOUR, 1, MINUTE, 2, |_, _| -1.0);
    let new = batch(T0 + 2 * DAY, 30, MINUTE, 4, |d, i| f64::from(d) + i as f64);
    let no_maintenance = AppendOptions { maintain: Some(false), ..Default::default() };
    let (t, _) = t.append(&[late], no_maintenance.clone()).await.unwrap();
    let (t, _) = t.append(&[new], no_maintenance).await.unwrap();
    check_answers(&lake).await;

    let s = Session::new(lake.clone());
    let refreshed = rows(&s, "REFRESH MATERIALIZED ROLLUP iot.r_1h").await;
    let text = arrow::util::pretty::pretty_format_batches(&refreshed).unwrap().to_string();
    // A changed hour is recomputed for every device: the late hour and the new one, 4 devices each.
    assert!(text.contains("| 8 "), "{text}");
    let bucket = "SELECT min_value FROM iot.r_1h WHERE device = 0 AND bucket = TIMESTAMP '2026-01-01 03:00:00'";
    assert_eq!(scalar_i64(&s, bucket).await, -1);
    check_answers(&lake).await;
    assert_eq!(rows(&s, "REFRESH MATERIALIZED ROLLUP iot.r_1h").await[0].num_rows(), 1);

    // Inline maintenance refreshes a rollup whose last refresh is older than its interval.
    let rollup = lake.table("iot.r_1h").await.unwrap();
    rollup.set_properties(&[("rollup-refresh-interval", Some("0s"))]).await.unwrap();
    let (_, result) = t
        .refresh()
        .await
        .unwrap()
        .append(&[batch(T0 + 2 * DAY + HOUR, 1, MINUTE, 1, |_, _| 7.0)], AppendOptions::default())
        .await
        .unwrap();
    let report = result.maintenance.unwrap().unwrap();
    assert_eq!((report.rollups_refreshed, report.rollup_rows), (1, 1));
    check_answers(&lake).await;
}

#[tokio::test(flavor = "multi_thread")]
async fn bad_definitions_and_drop() {
    let (_dir, lake) = lake_with_rollup().await;
    let s = Session::new(lake.clone());
    let create = |select: &str| format!("CREATE MATERIALIZED ROLLUP iot.bad AS {select}");
    for (select, msg) in [
        (
            "SELECT device, time_bucket(INTERVAL '1 hour', time) b, avg(value) FROM iot.r WHERE device = 1 GROUP BY 1, 2",
            "WHERE",
        ),
        ("SELECT site, time_bucket(INTERVAL '1 hour', time) b, avg(value) FROM iot.r GROUP BY 1, 2", "series-key"),
        ("SELECT device, time_bucket(INTERVAL '1 month', time) b, avg(value) FROM iot.r GROUP BY 1, 2", "time_bucket"),
        ("SELECT device, time_bucket(INTERVAL '1 hour', time) b, median(value) FROM iot.r GROUP BY 1, 2", "median"),
        ("SELECT device, avg(value) FROM iot.r GROUP BY 1", "time_bucket"),
        ("SELECT time_bucket(INTERVAL '1 hour', time) b, avg(value) FROM iot.r GROUP BY device, b", "grouping column"),
    ] {
        let err = s.sql(&create(select)).await.unwrap_err().to_string();
        assert!(err.contains(msg), "{select}: {err}");
    }
    assert!(lake.table("iot.bad").await.is_err(), "nothing is left behind");
    assert!(s.sql(CREATE).await.is_err(), "exists");
    rows(&s, &CREATE.replace("ROLLUP iot", "ROLLUP IF NOT EXISTS iot")).await;
    assert!(s.sql("DROP MATERIALIZED ROLLUP iot.r").await.unwrap_err().to_string().contains("not a rollup"));

    rows(&s, "DROP MATERIALIZED ROLLUP iot.r_1h").await;
    assert!(lake.table("iot.r_1h").await.is_err());
    let raw = lake.table("iot.r").await.unwrap();
    assert_eq!(raw.contract().rollups, Vec::<String>::new());
    assert!(!raw.metadata().properties().contains_key(property::ROLLUPS));
    assert!(s.sql("DROP MATERIALIZED ROLLUP iot.r_1h").await.is_err());
    rows(&s, "DROP MATERIALIZED ROLLUP IF EXISTS iot.r_1h").await;
}

#[tokio::test(flavor = "multi_thread")]
async fn a_rollback_of_the_raw_table_rebuilds_the_rollup() {
    let (_dir, lake) = lake_with_rollup().await;
    let t = lake.table("iot.r").await.unwrap();
    let before = t.metadata().current_snapshot_id().unwrap();
    let s = Session::new(lake.clone());
    t.append(&[batch(T0 + 2 * DAY, 60, MINUTE, 4, |_, _| 1.0)], AppendOptions::default()).await.unwrap();
    rows(&s, "REFRESH MATERIALIZED ROLLUP iot.r_1h").await;
    assert_eq!(scalar_i64(&s, "SELECT count(*) FROM iot.r_1h").await, 196);

    // The rollup now holds an hour the raw table no longer has: it isn't used until rebuilt.
    lake.table("iot.r").await.unwrap().rollback(before).await.unwrap();
    let sql = QUERIES[0].0;
    let unrouted = Session::new(lake.clone());
    unrouted.sql("SET timelake.rollup_routing = false").await.unwrap();
    assert!(!reads_rollup(&s, sql).await);
    assert_eq!(answer(&s, sql).await, answer(&unrouted, sql).await);

    rows(&s, "REFRESH MATERIALIZED ROLLUP iot.r_1h").await;
    assert_eq!(scalar_i64(&s, "SELECT count(*) FROM iot.r_1h").await, 192);
    check_answers(&lake).await;
}

#[tokio::test(flavor = "multi_thread")]
async fn rollups_keep_what_the_raw_table_drops() {
    let (_dir, lake) = lake_with_rollup().await;
    // The data is from January: a 30-day retention drops all of it from the raw table.
    let t = lake.table("iot.r").await.unwrap().set_properties(&[("retention", Some("30d"))]).await.unwrap();
    let (_, report) = t.maintain(timelake::MaintainOptions::default()).await.unwrap();
    assert!(report.partitions_dropped > 0);
    let s = Session::new(lake.clone());
    assert_eq!(scalar_i64(&s, "SELECT count(*) FROM iot.r").await, 0);
    assert_eq!(scalar_i64(&s, "SELECT count(*) FROM iot.r_1h").await, 192);
    let unrouted = Session::new(lake.clone());
    unrouted.sql("SET timelake.rollup_routing = false").await.unwrap();
    // Buckets older than the raw table's retention aren't read from the rollup.
    for &(sql, _) in QUERIES {
        assert_eq!(answer(&s, sql).await, answer(&unrouted, sql).await, "{sql}");
    }
}