Spaces:
Build error
Build error
File size: 14,838 Bytes
84d2a97 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 |
use std::collections::HashSet;
use std::sync::Arc;
use std::time::{Duration, Instant};
use futures::future::try_join_all;
use itertools::Itertools as _;
use rand::distributions::WeightedIndex;
use rand::rngs::StdRng;
use rand::{Rng, SeedableRng};
use segment::data_types::order_by::{Direction, OrderBy, OrderValue};
use segment::types::{
ExtendedPointId, Filter, ScoredPoint, WithPayload, WithPayloadInterface, WithVector,
};
use tokio::runtime::Handle;
use tokio::time::error::Elapsed;
use super::LocalShard;
use crate::collection_manager::holders::segment_holder::LockedSegment;
use crate::collection_manager::segments_searcher::SegmentsSearcher;
use crate::common::stopping_guard::StoppingGuard;
use crate::operations::types::{
CollectionError, CollectionResult, QueryScrollRequestInternal, RecordInternal, ScrollOrder,
};
impl LocalShard {
/// Basic parallel batching, it is conveniently used for the universal query API.
pub(super) async fn query_scroll_batch(
&self,
batch: Arc<Vec<QueryScrollRequestInternal>>,
search_runtime_handle: &Handle,
timeout: Duration,
) -> CollectionResult<Vec<Vec<ScoredPoint>>> {
let scrolls = batch
.iter()
.map(|request| self.query_scroll(request, search_runtime_handle, Some(timeout)));
// execute all the scrolls concurrently
let all_scroll_results = try_join_all(scrolls);
tokio::time::timeout(timeout, all_scroll_results)
.await
.map_err(|_| {
log::debug!(
"Query scroll timeout reached: {} seconds",
timeout.as_secs()
);
CollectionError::timeout(timeout.as_secs() as usize, "Query scroll")
})?
}
/// Scroll a single page, to be used for the universal query API only.
async fn query_scroll(
&self,
request: &QueryScrollRequestInternal,
search_runtime_handle: &Handle,
timeout: Option<Duration>,
) -> CollectionResult<Vec<ScoredPoint>> {
let QueryScrollRequestInternal {
limit,
with_vector,
filter,
scroll_order,
with_payload,
} = request;
let limit = *limit;
let offset_id = None;
let point_results = match scroll_order {
ScrollOrder::ById => self
.scroll_by_id(
offset_id,
limit,
with_payload,
with_vector,
filter.as_ref(),
search_runtime_handle,
timeout,
)
.await?
.into_iter()
.map(|record| ScoredPoint {
id: record.id,
version: 0,
score: 0.0,
payload: record.payload,
vector: record.vector,
shard_key: record.shard_key,
order_value: None,
})
.collect(),
ScrollOrder::ByField(order_by) => {
let (records, values) = self
.scroll_by_field(
limit,
with_payload,
with_vector,
filter.as_ref(),
search_runtime_handle,
order_by,
timeout,
)
.await?;
records
.into_iter()
.zip(values)
.map(|(record, value)| ScoredPoint {
id: record.id,
version: 0,
score: 0.0,
payload: record.payload,
vector: record.vector,
shard_key: record.shard_key,
order_value: Some(value),
})
.collect()
}
ScrollOrder::Random => {
let records = self
.scroll_randomly(
limit,
with_payload,
with_vector,
filter.as_ref(),
search_runtime_handle,
timeout,
)
.await?;
records
.into_iter()
.map(|record| ScoredPoint {
id: record.id,
version: 0,
score: 0.0,
payload: record.payload,
vector: record.vector,
shard_key: record.shard_key,
order_value: None,
})
.collect()
}
};
Ok(point_results)
}
#[allow(clippy::too_many_arguments)]
pub async fn scroll_by_id(
&self,
offset: Option<ExtendedPointId>,
limit: usize,
with_payload_interface: &WithPayloadInterface,
with_vector: &WithVector,
filter: Option<&Filter>,
search_runtime_handle: &Handle,
timeout: Option<Duration>,
) -> CollectionResult<Vec<RecordInternal>> {
let start = Instant::now();
let timeout = timeout.unwrap_or(self.shared_storage_config.search_timeout);
let stopping_guard = StoppingGuard::new();
let segments = self.segments.clone();
let (non_appendable, appendable) = segments.read().split_segments();
let read_filtered = |segment: LockedSegment| {
let filter = filter.cloned();
let is_stopped = stopping_guard.get_is_stopped();
search_runtime_handle.spawn_blocking(move || {
segment.get().read().read_filtered(
offset,
Some(limit),
filter.as_ref(),
&is_stopped,
)
})
};
let all_reads = tokio::time::timeout(
timeout,
try_join_all(
non_appendable
.into_iter()
.chain(appendable)
.map(read_filtered),
),
)
.await
.map_err(|_: Elapsed| {
CollectionError::timeout(timeout.as_secs() as usize, "scroll_by_id")
})??;
let point_ids = all_reads
.into_iter()
.flatten()
.sorted()
.dedup()
.take(limit)
.collect_vec();
let with_payload = WithPayload::from(with_payload_interface);
// update timeout
let timeout = timeout.saturating_sub(start.elapsed());
let mut records_map = tokio::time::timeout(
timeout,
SegmentsSearcher::retrieve(
segments,
&point_ids,
&with_payload,
with_vector,
search_runtime_handle,
),
)
.await
.map_err(|_: Elapsed| CollectionError::timeout(timeout.as_secs() as usize, "retrieve"))??;
let ordered_records = point_ids
.iter()
// Use remove to avoid cloning, we take each point ID only once
.filter_map(|point_id| records_map.remove(point_id))
.collect();
Ok(ordered_records)
}
#[allow(clippy::too_many_arguments)]
pub async fn scroll_by_field(
&self,
limit: usize,
with_payload_interface: &WithPayloadInterface,
with_vector: &WithVector,
filter: Option<&Filter>,
search_runtime_handle: &Handle,
order_by: &OrderBy,
timeout: Option<Duration>,
) -> CollectionResult<(Vec<RecordInternal>, Vec<OrderValue>)> {
let start = Instant::now();
let timeout = timeout.unwrap_or(self.shared_storage_config.search_timeout);
let stopping_guard = StoppingGuard::new();
let segments = self.segments.clone();
let (non_appendable, appendable) = segments.read().split_segments();
let read_ordered_filtered = |segment: LockedSegment| {
let is_stopped = stopping_guard.get_is_stopped();
let filter = filter.cloned();
let order_by = order_by.clone();
search_runtime_handle.spawn_blocking(move || {
segment.get().read().read_ordered_filtered(
Some(limit),
filter.as_ref(),
&order_by,
&is_stopped,
)
})
};
let all_reads = tokio::time::timeout(
timeout,
try_join_all(
non_appendable
.into_iter()
.chain(appendable)
.map(read_ordered_filtered),
),
)
.await
.map_err(|_: Elapsed| {
CollectionError::timeout(timeout.as_secs() as usize, "scroll_by_field")
})??;
let all_reads = all_reads.into_iter().collect::<Result<Vec<_>, _>>()?;
let (values, point_ids): (Vec<_>, Vec<_>) = all_reads
.into_iter()
.kmerge_by(|a, b| match order_by.direction() {
Direction::Asc => a <= b,
Direction::Desc => a >= b,
})
.dedup()
.take(limit)
.unzip();
let with_payload = WithPayload::from(with_payload_interface);
// update timeout
let timeout = timeout.saturating_sub(start.elapsed());
// Fetch with the requested vector and payload
let records_map = tokio::time::timeout(
timeout,
SegmentsSearcher::retrieve(
segments,
&point_ids,
&with_payload,
with_vector,
search_runtime_handle,
),
)
.await
.map_err(|_| CollectionError::timeout(timeout.as_secs() as usize, "retrieve"))??;
let ordered_records = point_ids
.iter()
.filter_map(|point_id| records_map.get(point_id).cloned())
.collect();
Ok((ordered_records, values))
}
async fn scroll_randomly(
&self,
limit: usize,
with_payload_interface: &WithPayloadInterface,
with_vector: &WithVector,
filter: Option<&Filter>,
search_runtime_handle: &Handle,
timeout: Option<Duration>,
) -> CollectionResult<Vec<RecordInternal>> {
let start = Instant::now();
let timeout = timeout.unwrap_or(self.shared_storage_config.search_timeout);
let stopping_guard = StoppingGuard::new();
let segments = self.segments.clone();
let (non_appendable, appendable) = segments.read().split_segments();
let read_filtered = |segment: LockedSegment| {
let is_stopped = stopping_guard.get_is_stopped();
let filter = filter.cloned();
search_runtime_handle.spawn_blocking(move || {
let get_segment = segment.get();
let read_segment = get_segment.read();
(
read_segment.available_point_count(),
read_segment.read_random_filtered(limit, filter.as_ref(), &is_stopped),
)
})
};
let all_reads = tokio::time::timeout(
timeout,
try_join_all(
non_appendable
.into_iter()
.chain(appendable)
.map(read_filtered),
),
)
.await
.map_err(|_: Elapsed| {
CollectionError::timeout(timeout.as_secs() as usize, "scroll_randomly")
})??;
let (availability, mut segments_reads): (Vec<_>, Vec<_>) = all_reads.into_iter().unzip();
// Shortcut if all segments are empty
if availability.iter().all(|&count| count == 0) {
return Ok(Vec::new());
}
// Select points in a weighted fashion from each segment, depending on how many points each segment has.
let distribution = WeightedIndex::new(availability).map_err(|err| {
CollectionError::service_error(format!(
"Failed to create weighted index for random scroll: {err:?}"
))
})?;
let mut rng = StdRng::from_entropy();
let mut random_points = HashSet::with_capacity(limit);
// Randomly sample points in two stages
//
// 1. This loop iterates <= LIMIT times, and either breaks early if we
// have enough points, or if some of the segments are exhausted.
//
// 2. If the segments are exhausted, we will fill up the rest of the
// points from other segments. In total, the complexity is guaranteed to
// be O(limit).
while random_points.len() < limit {
let segment_offset = rng.sample(&distribution);
let points = segments_reads.get_mut(segment_offset).unwrap();
if let Some(point) = points.pop() {
random_points.insert(point);
} else {
// It seems that some segments are empty early,
// so distribution does not make sense anymore.
// This is only possible if segments size < limit.
break;
}
}
// If we still need more points, we will get them from the rest of the segments.
// This is a rare case, as it seems we don't have enough points in individual segments.
// Therefore, we can ignore "proper" distribution, as it won't be accurate anyway.
if random_points.len() < limit {
let rest_points = segments_reads.into_iter().flatten();
for point in rest_points {
random_points.insert(point);
if random_points.len() >= limit {
break;
}
}
}
let selected_points: Vec<_> = random_points.into_iter().collect();
let with_payload = WithPayload::from(with_payload_interface);
// update timeout
let timeout = timeout.saturating_sub(start.elapsed());
let records_map = tokio::time::timeout(
timeout,
SegmentsSearcher::retrieve(
segments,
&selected_points,
&with_payload,
with_vector,
search_runtime_handle,
),
)
.await
.map_err(|_: Elapsed| CollectionError::timeout(timeout.as_secs() as usize, "retrieve"))??;
Ok(records_map.into_values().collect())
}
}
|