datafusion_expr_common/
accumulator.rs

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
// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements.  See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership.  The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License.  You may obtain a copy of the License at
//
//   http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing,
// software distributed under the License is distributed on an
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, either express or implied.  See the License for the
// specific language governing permissions and limitations
// under the License.

//! Accumulator module contains the trait definition for aggregation function's accumulators.

use arrow::array::ArrayRef;
use datafusion_common::{internal_err, Result, ScalarValue};
use std::fmt::Debug;

/// Tracks an aggregate function's state.
///
/// `Accumulator`s are stateful objects that implement a single group. They
/// aggregate values from multiple rows together into a final output aggregate.
///
/// [`GroupsAccumulator]` is an additional more performant (but also complex) API
/// that manages state for multiple groups at once.
///
/// An accumulator knows how to:
/// * update its state from inputs via [`update_batch`]
///
/// * compute the final value from its internal state via [`evaluate`]
///
/// * retract an update to its state from given inputs via
///   [`retract_batch`] (when used as a window aggregate [window
///   function])
///
/// * convert its internal state to a vector of aggregate values via
///   [`state`] and combine the state from multiple accumulators
///   via [`merge_batch`], as part of efficient multi-phase grouping.
///
/// [`GroupsAccumulator`]: crate::GroupsAccumulator
/// [`update_batch`]: Self::update_batch
/// [`retract_batch`]: Self::retract_batch
/// [`state`]: Self::state
/// [`evaluate`]: Self::evaluate
/// [`merge_batch`]: Self::merge_batch
/// [window function]: https://en.wikipedia.org/wiki/Window_function_(SQL)
pub trait Accumulator: Send + Sync + Debug {
    /// Updates the accumulator's state from its input.
    ///
    /// `values` contains the arguments to this aggregate function.
    ///
    /// For example, the `SUM` accumulator maintains a running sum,
    /// and `update_batch` adds each of the input values to the
    /// running sum.
    fn update_batch(&mut self, values: &[ArrayRef]) -> Result<()>;

    /// Returns the final aggregate value, consuming the internal state.
    ///
    /// For example, the `SUM` accumulator maintains a running sum,
    /// and `evaluate` will produce that running sum as its output.
    ///
    /// This function should not be called twice, otherwise it will
    /// result in potentially non-deterministic behavior.
    ///
    /// This function gets `&mut self` to allow for the accumulator to build
    /// arrow-compatible internal state that can be returned without copying
    /// when possible (for example distinct strings)
    fn evaluate(&mut self) -> Result<ScalarValue>;

    /// Returns the allocated size required for this accumulator, in
    /// bytes, including `Self`.
    ///
    /// This value is used to calculate the memory used during
    /// execution so DataFusion can stay within its allotted limit.
    ///
    /// "Allocated" means that for internal containers such as `Vec`,
    /// the `capacity` should be used not the `len`.
    fn size(&self) -> usize;

    /// Returns the intermediate state of the accumulator, consuming the
    /// intermediate state.
    ///
    /// This function should not be called twice, otherwise it will
    /// result in potentially non-deterministic behavior.
    ///
    /// This function gets `&mut self` to allow for the accumulator to build
    /// arrow-compatible internal state that can be returned without copying
    /// when possible (for example distinct strings).
    ///
    /// Intermediate state is used for "multi-phase" grouping in
    /// DataFusion, where an aggregate is computed in parallel with
    /// multiple `Accumulator` instances, as described below:
    ///
    /// # Multi-Phase Grouping
    ///
    /// ```text
    ///                               ▲
    ///                               │                   evaluate() is called to
    ///                               │                   produce the final aggregate
    ///                               │                   value per group
    ///                               │
    ///                  ┌─────────────────────────┐
    ///                  │GroupBy                  │
    ///                  │(AggregateMode::Final)   │      state() is called for each
    ///                  │                         │      group and the resulting
    ///                  └─────────────────────────┘      RecordBatches passed to the
    ///                               ▲
    ///                               │
    ///              ┌────────────────┴───────────────┐
    ///              │                                │
    ///              │                                │
    /// ┌─────────────────────────┐      ┌─────────────────────────┐
    /// │        GroubyBy         │      │        GroubyBy         │
    /// │(AggregateMode::Partial) │      │(AggregateMode::Partial) │
    /// └─────────────────────────┘      └─────────────────────────┘
    ///              ▲                                ▲
    ///              │                                │    update_batch() is called for
    ///              │                                │    each input RecordBatch
    ///         .─────────.                      .─────────.
    ///      ,─'           '─.                ,─'           '─.
    ///     ;      Input      :              ;      Input      :
    ///     :   Partition 0   ;              :   Partition 1   ;
    ///      ╲               ╱                ╲               ╱
    ///       '─.         ,─'                  '─.         ,─'
    ///          `───────'                        `───────'
    /// ```
    ///
    /// The partial state is serialized as `Arrays` and then combined
    /// with other partial states from different instances of this
    /// Accumulator (that ran on different partitions, for example).
    ///
    /// The state can be and often is a different type than the output
    /// type of the [`Accumulator`] and needs different merge
    /// operations (for example, the partial state for `COUNT` needs
    /// to be summed together)
    ///
    /// Some accumulators can return multiple values for their
    /// intermediate states. For example, the average accumulator
    /// tracks `sum` and `n`, and this function should return a vector
    /// of two values, sum and n.
    ///
    /// Note that [`ScalarValue::List`] can be used to pass multiple
    /// values if the number of intermediate values is not known at
    /// planning time (e.g. for `MEDIAN`)
    ///
    /// # Multi-phase repartitioned Grouping
    ///
    /// Many multi-phase grouping plans contain a Repartition operation
    /// as well as shown below:
    ///
    /// ```text
    ///                ▲                          ▲
    ///                │                          │
    ///                │                          │
    ///                │                          │
    ///                │                          │
    ///                │                          │
    ///    ┌───────────────────────┐  ┌───────────────────────┐       4. Each AggregateMode::Final
    ///    │GroupBy                │  │GroupBy                │       GroupBy has an entry for its
    ///    │(AggregateMode::Final) │  │(AggregateMode::Final) │       subset of groups (in this case
    ///    │                       │  │                       │       that means half the entries)
    ///    └───────────────────────┘  └───────────────────────┘
    ///                ▲                          ▲
    ///                │                          │
    ///                └─────────────┬────────────┘
    ///                              │
    ///                              │
    ///                              │
    ///                 ┌─────────────────────────┐                   3. Repartitioning by hash(group
    ///                 │       Repartition       │                   keys) ensures that each distinct
    ///                 │         HASH(x)         │                   group key now appears in exactly
    ///                 └─────────────────────────┘                   one partition
    ///                              ▲
    ///                              │
    ///              ┌───────────────┴─────────────┐
    ///              │                             │
    ///              │                             │
    /// ┌─────────────────────────┐  ┌──────────────────────────┐     2. Each AggregateMode::Partial
    /// │        GroubyBy         │  │         GroubyBy         │     GroupBy has an entry for *all*
    /// │(AggregateMode::Partial) │  │ (AggregateMode::Partial) │     the groups
    /// └─────────────────────────┘  └──────────────────────────┘
    ///              ▲                             ▲
    ///              │                             │
    ///              │                             │
    ///         .─────────.                   .─────────.
    ///      ,─'           '─.             ,─'           '─.
    ///     ;      Input      :           ;      Input      :         1. Since input data is
    ///     :   Partition 0   ;           :   Partition 1   ;         arbitrarily or RoundRobin
    ///      ╲               ╱             ╲               ╱          distributed, each partition
    ///       '─.         ,─'               '─.         ,─'           likely has all distinct
    ///          `───────'                     `───────'
    /// ```
    ///
    /// This structure is used so that the `AggregateMode::Partial` accumulators
    /// reduces the cardinality of the input as soon as possible. Typically,
    /// each partial accumulator sees all groups in the input as the group keys
    /// are evenly distributed across the input.
    ///
    /// The final output is computed by repartitioning the result of
    /// [`Self::state`] from each Partial aggregate and `hash(group keys)` so
    /// that each distinct group key appears in exactly one of the
    /// `AggregateMode::Final` GroupBy nodes. The outputs of the final nodes are
    /// then unioned together to produce the overall final output.
    ///
    /// Here is an example that shows the distribution of groups in the
    /// different phases
    ///
    /// ```text
    ///               ┌─────┐                ┌─────┐
    ///               │  1  │                │  3  │
    ///               ├─────┤                ├─────┤
    ///               │  2  │                │  4  │                After repartitioning by
    ///               └─────┘                └─────┘                hash(group keys), each distinct
    ///               ┌─────┐                ┌─────┐                group key now appears in exactly
    ///               │  1  │                │  3  │                one partition
    ///               ├─────┤                ├─────┤
    ///               │  2  │                │  4  │
    ///               └─────┘                └─────┘
    ///
    ///
    /// ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─
    ///
    ///               ┌─────┐                ┌─────┐
    ///               │  2  │                │  2  │
    ///               ├─────┤                ├─────┤
    ///               │  1  │                │  2  │
    ///               ├─────┤                ├─────┤
    ///               │  3  │                │  3  │
    ///               ├─────┤                ├─────┤
    ///               │  4  │                │  1  │
    ///               └─────┘                └─────┘                Input data is arbitrarily or
    ///                 ...                    ...                  RoundRobin distributed, each
    ///               ┌─────┐                ┌─────┐                partition likely has all
    ///               │  1  │                │  4  │                distinct group keys
    ///               ├─────┤                ├─────┤
    ///               │  4  │                │  3  │
    ///               ├─────┤                ├─────┤
    ///               │  1  │                │  1  │
    ///               ├─────┤                ├─────┤
    ///               │  4  │                │  3  │
    ///               └─────┘                └─────┘
    ///
    ///           group values           group values
    ///           in partition 0         in partition 1
    /// ```
    fn state(&mut self) -> Result<Vec<ScalarValue>>;

    /// Updates the accumulator's state from an `Array` containing one
    /// or more intermediate values.
    ///
    /// For some aggregates (such as `SUM`), merge_batch is the same
    /// as `update_batch`, but for some aggregrates (such as `COUNT`)
    /// the operations differ. See [`Self::state`] for more details on how
    /// state is used and merged.
    ///
    /// The `states` array passed was formed by concatenating the
    /// results of calling [`Self::state`] on zero or more other
    /// `Accumulator` instances.
    fn merge_batch(&mut self, states: &[ArrayRef]) -> Result<()>;

    /// Retracts (removed) an update (caused by the given inputs) to
    /// accumulator's state.
    ///
    /// This is the inverse operation of [`Self::update_batch`] and is used
    /// to incrementally calculate window aggregates where the `OVER`
    /// clause defines a bounded window.
    ///
    /// # Example
    ///
    /// For example, given the following input partition
    ///
    /// ```text
    ///                     │      current      │
    ///                            window
    ///                     │                   │
    ///                ┌────┬────┬────┬────┬────┬────┬────┬────┬────┐
    ///     Input      │ A  │ B  │ C  │ D  │ E  │ F  │ G  │ H  │ I  │
    ///   partition    └────┴────┴────┴────┼────┴────┴────┴────┼────┘
    ///
    ///                                    │         next      │
    ///                                             window
    /// ```
    ///
    /// First, [`Self::evaluate`] will be called to produce the output
    /// for the current window.
    ///
    /// Then, to advance to the next window:
    ///
    /// First, [`Self::retract_batch`] will be called with the values
    /// that are leaving the window, `[B, C, D]` and then
    /// [`Self::update_batch`] will be called with the values that are
    /// entering the window, `[F, G, H]`.
    fn retract_batch(&mut self, _values: &[ArrayRef]) -> Result<()> {
        // TODO add retract for all accumulators
        internal_err!(
            "Retract should be implemented for aggregate functions when used with custom window frame queries"
        )
    }

    /// Does the accumulator support incrementally updating its value
    /// by *removing* values.
    ///
    /// If this function returns true, [`Self::retract_batch`] will be
    /// called for sliding window functions such as queries with an
    /// `OVER (ROWS BETWEEN 1 PRECEDING AND 2 FOLLOWING)`
    fn supports_retract_batch(&self) -> bool {
        false
    }
}