Line data Source code
1 : use std::{sync::Arc, time::SystemTime};
2 :
3 : use anyhow::Context;
4 : use bytes::{buf::Writer, BufMut, BytesMut};
5 : use chrono::{Datelike, Timelike};
6 : use futures::{Stream, StreamExt};
7 : use parquet::{
8 : basic::Compression,
9 : file::{
10 : metadata::RowGroupMetaDataPtr,
11 : properties::{WriterProperties, WriterPropertiesPtr, DEFAULT_PAGE_SIZE},
12 : writer::SerializedFileWriter,
13 : },
14 : record::RecordWriter,
15 : };
16 : use pq_proto::StartupMessageParams;
17 : use remote_storage::{GenericRemoteStorage, RemotePath, RemoteStorageConfig, TimeoutOrCancel};
18 : use serde::ser::SerializeMap;
19 : use tokio::{sync::mpsc, time};
20 : use tokio_util::sync::CancellationToken;
21 : use tracing::{debug, info, Span};
22 : use utils::backoff;
23 :
24 : use crate::{config::remote_storage_from_toml, context::LOG_CHAN_DISCONNECT};
25 :
26 : use super::{RequestMonitoring, LOG_CHAN};
27 :
28 12 : #[derive(clap::Args, Clone, Debug)]
29 : pub struct ParquetUploadArgs {
30 : /// Storage location to upload the parquet files to.
31 : /// Encoded as toml (same format as pageservers), eg
32 : /// `{bucket_name='the-bucket',bucket_region='us-east-1',prefix_in_bucket='proxy',endpoint='http://minio:9000'}`
33 : #[clap(long, value_parser = remote_storage_from_toml)]
34 : parquet_upload_remote_storage: Option<RemoteStorageConfig>,
35 :
36 : #[clap(long, value_parser = remote_storage_from_toml)]
37 : parquet_upload_disconnect_events_remote_storage: Option<RemoteStorageConfig>,
38 :
39 : /// How many rows to include in a row group
40 6 : #[clap(long, default_value_t = 8192)]
41 0 : parquet_upload_row_group_size: usize,
42 :
43 : /// How large each column page should be in bytes
44 6 : #[clap(long, default_value_t = DEFAULT_PAGE_SIZE)]
45 0 : parquet_upload_page_size: usize,
46 :
47 : /// How large the total parquet file should be in bytes
48 6 : #[clap(long, default_value_t = 100_000_000)]
49 0 : parquet_upload_size: i64,
50 :
51 : /// How long to wait before forcing a file upload
52 : #[clap(long, default_value = "20m", value_parser = humantime::parse_duration)]
53 0 : parquet_upload_maximum_duration: tokio::time::Duration,
54 :
55 : /// What level of compression to use
56 6 : #[clap(long, default_value_t = Compression::UNCOMPRESSED)]
57 0 : parquet_upload_compression: Compression,
58 : }
59 :
60 : // Occasional network issues and such can cause remote operations to fail, and
61 : // that's expected. If a upload fails, we log it at info-level, and retry.
62 : // But after FAILED_UPLOAD_WARN_THRESHOLD retries, we start to log it at WARN
63 : // level instead, as repeated failures can mean a more serious problem. If it
64 : // fails more than FAILED_UPLOAD_RETRIES times, we give up
65 : pub const FAILED_UPLOAD_WARN_THRESHOLD: u32 = 3;
66 : pub const FAILED_UPLOAD_MAX_RETRIES: u32 = 10;
67 :
68 : // the parquet crate leaves a lot to be desired...
69 : // what follows is an attempt to write parquet files with minimal allocs.
70 : // complication: parquet is a columnar format, while we want to write in as rows.
71 : // design:
72 : // * we batch up to 1024 rows, then flush them into a 'row group'
73 : // * after each rowgroup write, we check the length of the file and upload to s3 if large enough
74 :
75 8360222 : #[derive(parquet_derive::ParquetRecordWriter)]
76 : pub struct RequestData {
77 : region: &'static str,
78 : protocol: &'static str,
79 : /// Must be UTC. The derive macro doesn't like the timezones
80 : timestamp: chrono::NaiveDateTime,
81 : session_id: uuid::Uuid,
82 : peer_addr: String,
83 : username: Option<String>,
84 : application_name: Option<String>,
85 : endpoint_id: Option<String>,
86 : database: Option<String>,
87 : project: Option<String>,
88 : branch: Option<String>,
89 : pg_options: Option<String>,
90 : auth_method: Option<&'static str>,
91 : error: Option<&'static str>,
92 : /// Success is counted if we form a HTTP response with sql rows inside
93 : /// Or if we make it to proxy_pass
94 : success: bool,
95 : /// Indicates if the cplane started the new compute node for this request.
96 : cold_start_info: &'static str,
97 : /// Tracks time from session start (HTTP request/libpq TCP handshake)
98 : /// Through to success/failure
99 : duration_us: u64,
100 : /// If the session was successful after the disconnect, will be created one more event with filled `disconnect_timestamp`.
101 : disconnect_timestamp: Option<chrono::NaiveDateTime>,
102 : }
103 :
104 : struct Options<'a> {
105 : options: &'a StartupMessageParams,
106 : }
107 :
108 : impl<'a> serde::Serialize for Options<'a> {
109 0 : fn serialize<S>(&self, s: S) -> Result<S::Ok, S::Error>
110 0 : where
111 0 : S: serde::Serializer,
112 0 : {
113 0 : let mut state = s.serialize_map(None)?;
114 0 : for (k, v) in self.options.iter() {
115 0 : state.serialize_entry(k, v)?;
116 : }
117 0 : state.end()
118 0 : }
119 : }
120 :
121 : impl From<&RequestMonitoring> for RequestData {
122 0 : fn from(value: &RequestMonitoring) -> Self {
123 0 : Self {
124 0 : session_id: value.session_id,
125 0 : peer_addr: value.peer_addr.to_string(),
126 0 : timestamp: value.first_packet.naive_utc(),
127 0 : username: value.user.as_deref().map(String::from),
128 0 : application_name: value.application.as_deref().map(String::from),
129 0 : endpoint_id: value.endpoint_id.as_deref().map(String::from),
130 0 : database: value.dbname.as_deref().map(String::from),
131 0 : project: value.project.as_deref().map(String::from),
132 0 : branch: value.branch.as_deref().map(String::from),
133 0 : pg_options: value
134 0 : .pg_options
135 0 : .as_ref()
136 0 : .and_then(|options| serde_json::to_string(&Options { options }).ok()),
137 0 : auth_method: value.auth_method.as_ref().map(|x| match x {
138 0 : super::AuthMethod::Web => "web",
139 0 : super::AuthMethod::ScramSha256 => "scram_sha_256",
140 0 : super::AuthMethod::ScramSha256Plus => "scram_sha_256_plus",
141 0 : super::AuthMethod::Cleartext => "cleartext",
142 0 : }),
143 0 : protocol: value.protocol.as_str(),
144 0 : region: value.region,
145 0 : error: value.error_kind.as_ref().map(|e| e.to_metric_label()),
146 0 : success: value.success,
147 0 : cold_start_info: value.cold_start_info.as_str(),
148 0 : duration_us: SystemTime::from(value.first_packet)
149 0 : .elapsed()
150 0 : .unwrap_or_default()
151 0 : .as_micros() as u64, // 584 millenia... good enough
152 0 : disconnect_timestamp: value.disconnect_timestamp.map(|x| x.naive_utc()),
153 0 : }
154 0 : }
155 : }
156 :
157 : /// Parquet request context worker
158 : ///
159 : /// It listened on a channel for all completed requests, extracts the data and writes it into a parquet file,
160 : /// then uploads a completed batch to S3
161 0 : pub async fn worker(
162 0 : cancellation_token: CancellationToken,
163 0 : config: ParquetUploadArgs,
164 0 : ) -> anyhow::Result<()> {
165 0 : let Some(remote_storage_config) = config.parquet_upload_remote_storage else {
166 0 : tracing::warn!("parquet request upload: no s3 bucket configured");
167 0 : return Ok(());
168 : };
169 :
170 0 : let (tx, mut rx) = mpsc::unbounded_channel();
171 0 : LOG_CHAN.set(tx.downgrade()).unwrap();
172 0 :
173 0 : // setup row stream that will close on cancellation
174 0 : let cancellation_token2 = cancellation_token.clone();
175 0 : tokio::spawn(async move {
176 0 : cancellation_token2.cancelled().await;
177 : // dropping this sender will cause the channel to close only once
178 : // all the remaining inflight requests have been completed.
179 0 : drop(tx);
180 0 : });
181 0 : let rx = futures::stream::poll_fn(move |cx| rx.poll_recv(cx));
182 0 : let rx = rx.map(RequestData::from);
183 :
184 0 : let storage =
185 0 : GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?;
186 :
187 0 : let properties = WriterProperties::builder()
188 0 : .set_data_page_size_limit(config.parquet_upload_page_size)
189 0 : .set_compression(config.parquet_upload_compression);
190 0 :
191 0 : let parquet_config = ParquetConfig {
192 0 : propeties: Arc::new(properties.build()),
193 0 : rows_per_group: config.parquet_upload_row_group_size,
194 0 : file_size: config.parquet_upload_size,
195 0 : max_duration: config.parquet_upload_maximum_duration,
196 0 :
197 0 : #[cfg(any(test, feature = "testing"))]
198 0 : test_remote_failures: 0,
199 0 : };
200 :
201 : // TODO(anna): consider moving this to a separate function.
202 0 : if let Some(disconnect_events_storage_config) =
203 0 : config.parquet_upload_disconnect_events_remote_storage
204 : {
205 0 : let (tx_disconnect, mut rx_disconnect) = mpsc::unbounded_channel();
206 0 : LOG_CHAN_DISCONNECT.set(tx_disconnect.downgrade()).unwrap();
207 0 :
208 0 : // setup row stream that will close on cancellation
209 0 : tokio::spawn(async move {
210 0 : cancellation_token.cancelled().await;
211 : // dropping this sender will cause the channel to close only once
212 : // all the remaining inflight requests have been completed.
213 0 : drop(tx_disconnect);
214 0 : });
215 0 : let rx_disconnect = futures::stream::poll_fn(move |cx| rx_disconnect.poll_recv(cx));
216 0 : let rx_disconnect = rx_disconnect.map(RequestData::from);
217 :
218 0 : let storage_disconnect =
219 0 : GenericRemoteStorage::from_config(&disconnect_events_storage_config)
220 0 : .context("remote storage for disconnect events init")?;
221 0 : let parquet_config_disconnect = parquet_config.clone();
222 : tokio::try_join!(
223 : worker_inner(storage, rx, parquet_config),
224 : worker_inner(storage_disconnect, rx_disconnect, parquet_config_disconnect)
225 : )
226 0 : .map(|_| ())
227 : } else {
228 0 : worker_inner(storage, rx, parquet_config).await
229 : }
230 0 : }
231 :
232 : #[derive(Clone, Debug)]
233 : struct ParquetConfig {
234 : propeties: WriterPropertiesPtr,
235 : rows_per_group: usize,
236 : file_size: i64,
237 :
238 : max_duration: tokio::time::Duration,
239 :
240 : #[cfg(any(test, feature = "testing"))]
241 : test_remote_failures: u64,
242 : }
243 :
244 10 : async fn worker_inner(
245 10 : storage: GenericRemoteStorage,
246 10 : rx: impl Stream<Item = RequestData>,
247 10 : config: ParquetConfig,
248 10 : ) -> anyhow::Result<()> {
249 : #[cfg(any(test, feature = "testing"))]
250 10 : let storage = if config.test_remote_failures > 0 {
251 4 : GenericRemoteStorage::unreliable_wrapper(storage, config.test_remote_failures)
252 : } else {
253 6 : storage
254 : };
255 :
256 10 : let mut rx = std::pin::pin!(rx);
257 10 :
258 10 : let mut rows = Vec::with_capacity(config.rows_per_group);
259 :
260 10 : let schema = rows.as_slice().schema()?;
261 10 : let buffer = BytesMut::new();
262 10 : let w = buffer.writer();
263 10 : let mut w = SerializedFileWriter::new(w, schema.clone(), config.propeties.clone())?;
264 :
265 10 : let mut last_upload = time::Instant::now();
266 10 :
267 10 : let mut len = 0;
268 418010 : while let Some(row) = rx.next().await {
269 418000 : rows.push(row);
270 418000 : let force = last_upload.elapsed() > config.max_duration;
271 418000 : if rows.len() == config.rows_per_group || force {
272 210 : let rg_meta;
273 210 : (rows, w, rg_meta) = flush_rows(rows, w).await?;
274 210 : len += rg_meta.compressed_size();
275 417790 : }
276 418000 : if len > config.file_size || force {
277 56 : last_upload = time::Instant::now();
278 234 : let file = upload_parquet(w, len, &storage).await?;
279 56 : w = SerializedFileWriter::new(file, schema.clone(), config.propeties.clone())?;
280 56 : len = 0;
281 417944 : }
282 : }
283 :
284 10 : if !rows.is_empty() {
285 2 : let rg_meta;
286 2 : (_, w, rg_meta) = flush_rows(rows, w).await?;
287 2 : len += rg_meta.compressed_size();
288 8 : }
289 :
290 10 : if !w.flushed_row_groups().is_empty() {
291 24 : let _: Writer<BytesMut> = upload_parquet(w, len, &storage).await?;
292 4 : }
293 :
294 10 : Ok(())
295 10 : }
296 :
297 212 : async fn flush_rows<W>(
298 212 : rows: Vec<RequestData>,
299 212 : mut w: SerializedFileWriter<W>,
300 212 : ) -> anyhow::Result<(
301 212 : Vec<RequestData>,
302 212 : SerializedFileWriter<W>,
303 212 : RowGroupMetaDataPtr,
304 212 : )>
305 212 : where
306 212 : W: std::io::Write + Send + 'static,
307 212 : {
308 212 : let span = Span::current();
309 212 : let (mut rows, w, rg_meta) = tokio::task::spawn_blocking(move || {
310 212 : let _enter = span.enter();
311 :
312 212 : let mut rg = w.next_row_group()?;
313 212 : rows.as_slice().write_to_row_group(&mut rg)?;
314 212 : let rg_meta = rg.close()?;
315 :
316 212 : let size = rg_meta.compressed_size();
317 212 : let compression = rg_meta.compressed_size() as f64 / rg_meta.total_byte_size() as f64;
318 212 :
319 212 : debug!(size, compression, "flushed row group to parquet file");
320 :
321 212 : Ok::<_, parquet::errors::ParquetError>((rows, w, rg_meta))
322 212 : })
323 212 : .await
324 212 : .unwrap()?;
325 :
326 212 : rows.clear();
327 212 : Ok((rows, w, rg_meta))
328 212 : }
329 :
330 62 : async fn upload_parquet(
331 62 : mut w: SerializedFileWriter<Writer<BytesMut>>,
332 62 : len: i64,
333 62 : storage: &GenericRemoteStorage,
334 62 : ) -> anyhow::Result<Writer<BytesMut>> {
335 62 : let len_uncompressed = w
336 62 : .flushed_row_groups()
337 62 : .iter()
338 212 : .map(|rg| rg.total_byte_size())
339 62 : .sum::<i64>();
340 :
341 : // I don't know how compute intensive this is, although it probably isn't much... better be safe than sorry.
342 : // finish method only available on the fork: https://github.com/apache/arrow-rs/issues/5253
343 62 : let (mut buffer, metadata) =
344 62 : tokio::task::spawn_blocking(move || -> parquet::errors::Result<_> {
345 62 : let metadata = w.finish()?;
346 62 : let buffer = std::mem::take(w.inner_mut().get_mut());
347 62 : Ok((buffer, metadata))
348 62 : })
349 62 : .await
350 62 : .unwrap()?;
351 :
352 62 : let data = buffer.split().freeze();
353 62 :
354 62 : let compression = len as f64 / len_uncompressed as f64;
355 62 : let size = data.len();
356 62 : let now = chrono::Utc::now();
357 62 : let id = uuid::Uuid::new_v7(uuid::Timestamp::from_unix(
358 62 : uuid::NoContext,
359 62 : // we won't be running this in 1970. this cast is ok
360 62 : now.timestamp() as u64,
361 62 : now.timestamp_subsec_nanos(),
362 62 : ));
363 62 :
364 62 : info!(
365 : %id,
366 : rows = metadata.num_rows,
367 0 : size, compression, "uploading request parquet file"
368 : );
369 :
370 62 : let year = now.year();
371 62 : let month = now.month();
372 62 : let day = now.day();
373 62 : let hour = now.hour();
374 : // segment files by time for S3 performance
375 62 : let path = RemotePath::from_string(&format!(
376 62 : "{year:04}/{month:02}/{day:02}/{hour:02}/requests_{id}.parquet"
377 62 : ))?;
378 62 : let cancel = CancellationToken::new();
379 62 : let maybe_err = backoff::retry(
380 86 : || async {
381 86 : let stream = futures::stream::once(futures::future::ready(Ok(data.clone())));
382 86 : storage
383 86 : .upload(stream, data.len(), &path, None, &cancel)
384 196 : .await
385 86 : },
386 62 : TimeoutOrCancel::caused_by_cancel,
387 62 : FAILED_UPLOAD_WARN_THRESHOLD,
388 62 : FAILED_UPLOAD_MAX_RETRIES,
389 62 : "request_data_upload",
390 62 : // we don't want cancellation to interrupt here, so we make a dummy cancel token
391 62 : &cancel,
392 62 : )
393 196 : .await
394 62 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
395 62 : .and_then(|x| x)
396 62 : .context("request_data_upload")
397 62 : .err();
398 :
399 62 : if let Some(err) = maybe_err {
400 0 : tracing::warn!(%id, %err, "failed to upload request data");
401 62 : }
402 :
403 62 : Ok(buffer.writer())
404 62 : }
405 :
406 : #[cfg(test)]
407 : mod tests {
408 : use std::{net::Ipv4Addr, num::NonZeroUsize, sync::Arc};
409 :
410 : use camino::Utf8Path;
411 : use clap::Parser;
412 : use futures::{Stream, StreamExt};
413 : use itertools::Itertools;
414 : use parquet::{
415 : basic::{Compression, ZstdLevel},
416 : file::{
417 : properties::{WriterProperties, DEFAULT_PAGE_SIZE},
418 : reader::FileReader,
419 : serialized_reader::SerializedFileReader,
420 : },
421 : };
422 : use rand::{rngs::StdRng, Rng, SeedableRng};
423 : use remote_storage::{
424 : GenericRemoteStorage, RemoteStorageConfig, RemoteStorageKind, S3Config,
425 : DEFAULT_MAX_KEYS_PER_LIST_RESPONSE, DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT,
426 : };
427 : use tokio::{sync::mpsc, time};
428 : use walkdir::WalkDir;
429 :
430 : use super::{worker_inner, ParquetConfig, ParquetUploadArgs, RequestData};
431 :
432 4 : #[derive(Parser)]
433 : struct ProxyCliArgs {
434 : #[clap(flatten)]
435 : parquet_upload: ParquetUploadArgs,
436 : }
437 :
438 : #[test]
439 2 : fn default_parser() {
440 2 : let ProxyCliArgs { parquet_upload } = ProxyCliArgs::parse_from(["proxy"]);
441 2 : assert_eq!(parquet_upload.parquet_upload_remote_storage, None);
442 2 : assert_eq!(parquet_upload.parquet_upload_row_group_size, 8192);
443 2 : assert_eq!(parquet_upload.parquet_upload_page_size, DEFAULT_PAGE_SIZE);
444 2 : assert_eq!(parquet_upload.parquet_upload_size, 100_000_000);
445 2 : assert_eq!(
446 2 : parquet_upload.parquet_upload_maximum_duration,
447 2 : time::Duration::from_secs(20 * 60)
448 2 : );
449 2 : assert_eq!(
450 2 : parquet_upload.parquet_upload_compression,
451 2 : Compression::UNCOMPRESSED
452 2 : );
453 2 : }
454 :
455 : #[test]
456 2 : fn full_parser() {
457 2 : let ProxyCliArgs { parquet_upload } = ProxyCliArgs::parse_from([
458 2 : "proxy",
459 2 : "--parquet-upload-remote-storage",
460 2 : "{bucket_name='default',prefix_in_bucket='proxy/',bucket_region='us-east-1',endpoint='http://minio:9000'}",
461 2 : "--parquet-upload-row-group-size",
462 2 : "100",
463 2 : "--parquet-upload-page-size",
464 2 : "10000",
465 2 : "--parquet-upload-size",
466 2 : "10000000",
467 2 : "--parquet-upload-maximum-duration",
468 2 : "10m",
469 2 : "--parquet-upload-compression",
470 2 : "zstd(5)",
471 2 : ]);
472 2 : assert_eq!(
473 2 : parquet_upload.parquet_upload_remote_storage,
474 2 : Some(RemoteStorageConfig {
475 2 : storage: RemoteStorageKind::AwsS3(S3Config {
476 2 : bucket_name: "default".into(),
477 2 : bucket_region: "us-east-1".into(),
478 2 : prefix_in_bucket: Some("proxy/".into()),
479 2 : endpoint: Some("http://minio:9000".into()),
480 2 : concurrency_limit: NonZeroUsize::new(
481 2 : DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT
482 2 : )
483 2 : .unwrap(),
484 2 : max_keys_per_list_response: DEFAULT_MAX_KEYS_PER_LIST_RESPONSE,
485 2 : upload_storage_class: None,
486 2 : }),
487 2 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
488 2 : })
489 2 : );
490 2 : assert_eq!(parquet_upload.parquet_upload_row_group_size, 100);
491 2 : assert_eq!(parquet_upload.parquet_upload_page_size, 10000);
492 2 : assert_eq!(parquet_upload.parquet_upload_size, 10_000_000);
493 2 : assert_eq!(
494 2 : parquet_upload.parquet_upload_maximum_duration,
495 2 : time::Duration::from_secs(10 * 60)
496 2 : );
497 2 : assert_eq!(
498 2 : parquet_upload.parquet_upload_compression,
499 2 : Compression::ZSTD(ZstdLevel::try_new(5).unwrap())
500 2 : );
501 2 : }
502 :
503 418000 : fn generate_request_data(rng: &mut impl Rng) -> RequestData {
504 418000 : RequestData {
505 418000 : session_id: uuid::Builder::from_random_bytes(rng.gen()).into_uuid(),
506 418000 : peer_addr: Ipv4Addr::from(rng.gen::<[u8; 4]>()).to_string(),
507 418000 : timestamp: chrono::DateTime::from_timestamp_millis(
508 418000 : rng.gen_range(1703862754..1803862754),
509 418000 : )
510 418000 : .unwrap()
511 418000 : .naive_utc(),
512 418000 : application_name: Some("test".to_owned()),
513 418000 : username: Some(hex::encode(rng.gen::<[u8; 4]>())),
514 418000 : endpoint_id: Some(hex::encode(rng.gen::<[u8; 16]>())),
515 418000 : database: Some(hex::encode(rng.gen::<[u8; 16]>())),
516 418000 : project: Some(hex::encode(rng.gen::<[u8; 16]>())),
517 418000 : branch: Some(hex::encode(rng.gen::<[u8; 16]>())),
518 418000 : pg_options: None,
519 418000 : auth_method: None,
520 418000 : protocol: ["tcp", "ws", "http"][rng.gen_range(0..3)],
521 418000 : region: "us-east-1",
522 418000 : error: None,
523 418000 : success: rng.gen(),
524 418000 : cold_start_info: "no",
525 418000 : duration_us: rng.gen_range(0..30_000_000),
526 418000 : disconnect_timestamp: None,
527 418000 : }
528 418000 : }
529 :
530 14 : fn random_stream(len: usize) -> impl Stream<Item = RequestData> + Unpin {
531 14 : let mut rng = StdRng::from_seed([0x39; 32]);
532 14 : futures::stream::iter(
533 418000 : std::iter::repeat_with(move || generate_request_data(&mut rng)).take(len),
534 14 : )
535 14 : }
536 :
537 10 : async fn run_test(
538 10 : tmpdir: &Utf8Path,
539 10 : config: ParquetConfig,
540 10 : rx: impl Stream<Item = RequestData>,
541 10 : ) -> Vec<(u64, usize, i64)> {
542 10 : let remote_storage_config = RemoteStorageConfig {
543 10 : storage: RemoteStorageKind::LocalFs {
544 10 : local_path: tmpdir.to_path_buf(),
545 10 : },
546 10 : timeout: std::time::Duration::from_secs(120),
547 10 : };
548 10 : let storage = GenericRemoteStorage::from_config(&remote_storage_config).unwrap();
549 10 :
550 610 : worker_inner(storage, rx, config).await.unwrap();
551 10 :
552 10 : let mut files = WalkDir::new(tmpdir.as_std_path())
553 10 : .into_iter()
554 112 : .filter_map(|entry| entry.ok())
555 112 : .filter(|entry| entry.file_type().is_file())
556 62 : .map(|entry| entry.path().to_path_buf())
557 10 : .collect_vec();
558 10 : files.sort();
559 10 :
560 10 : files
561 10 : .into_iter()
562 62 : .map(|path| std::fs::File::open(tmpdir.as_std_path().join(path)).unwrap())
563 62 : .map(|file| {
564 62 : (
565 62 : file.metadata().unwrap(),
566 62 : SerializedFileReader::new(file).unwrap().metadata().clone(),
567 62 : )
568 62 : })
569 62 : .map(|(file_meta, parquet_meta)| {
570 62 : (
571 62 : file_meta.len(),
572 62 : parquet_meta.num_row_groups(),
573 62 : parquet_meta.file_metadata().num_rows(),
574 62 : )
575 62 : })
576 10 : .collect()
577 10 : }
578 :
579 : #[tokio::test]
580 2 : async fn verify_parquet_no_compression() {
581 2 : let tmpdir = camino_tempfile::tempdir().unwrap();
582 2 :
583 2 : let config = ParquetConfig {
584 2 : propeties: Arc::new(WriterProperties::new()),
585 2 : rows_per_group: 2_000,
586 2 : file_size: 1_000_000,
587 2 : max_duration: time::Duration::from_secs(20 * 60),
588 2 : test_remote_failures: 0,
589 2 : };
590 2 :
591 2 : let rx = random_stream(50_000);
592 124 : let file_stats = run_test(tmpdir.path(), config, rx).await;
593 2 :
594 2 : assert_eq!(
595 2 : file_stats,
596 2 : [
597 2 : (1315874, 3, 6000),
598 2 : (1315867, 3, 6000),
599 2 : (1315927, 3, 6000),
600 2 : (1315884, 3, 6000),
601 2 : (1316014, 3, 6000),
602 2 : (1315856, 3, 6000),
603 2 : (1315648, 3, 6000),
604 2 : (1315884, 3, 6000),
605 2 : (438913, 1, 2000)
606 2 : ]
607 2 : );
608 2 :
609 2 : tmpdir.close().unwrap();
610 2 : }
611 :
612 : #[tokio::test]
613 2 : async fn verify_parquet_min_compression() {
614 2 : let tmpdir = camino_tempfile::tempdir().unwrap();
615 2 :
616 2 : let config = ParquetConfig {
617 2 : propeties: Arc::new(
618 2 : WriterProperties::builder()
619 2 : .set_compression(parquet::basic::Compression::ZSTD(ZstdLevel::default()))
620 2 : .build(),
621 2 : ),
622 2 : rows_per_group: 2_000,
623 2 : file_size: 1_000_000,
624 2 : max_duration: time::Duration::from_secs(20 * 60),
625 2 : test_remote_failures: 0,
626 2 : };
627 2 :
628 2 : let rx = random_stream(50_000);
629 92 : let file_stats = run_test(tmpdir.path(), config, rx).await;
630 2 :
631 2 : // with compression, there are fewer files with more rows per file
632 2 : assert_eq!(
633 2 : file_stats,
634 2 : [
635 2 : (1223214, 5, 10000),
636 2 : (1229364, 5, 10000),
637 2 : (1231158, 5, 10000),
638 2 : (1230520, 5, 10000),
639 2 : (1221798, 5, 10000)
640 2 : ]
641 2 : );
642 2 :
643 2 : tmpdir.close().unwrap();
644 2 : }
645 :
646 : #[tokio::test]
647 2 : async fn verify_parquet_strong_compression() {
648 2 : let tmpdir = camino_tempfile::tempdir().unwrap();
649 2 :
650 2 : let config = ParquetConfig {
651 2 : propeties: Arc::new(
652 2 : WriterProperties::builder()
653 2 : .set_compression(parquet::basic::Compression::ZSTD(
654 2 : ZstdLevel::try_new(10).unwrap(),
655 2 : ))
656 2 : .build(),
657 2 : ),
658 2 : rows_per_group: 2_000,
659 2 : file_size: 1_000_000,
660 2 : max_duration: time::Duration::from_secs(20 * 60),
661 2 : test_remote_failures: 0,
662 2 : };
663 2 :
664 2 : let rx = random_stream(50_000);
665 92 : let file_stats = run_test(tmpdir.path(), config, rx).await;
666 2 :
667 2 : // with strong compression, the files are smaller
668 2 : assert_eq!(
669 2 : file_stats,
670 2 : [
671 2 : (1208861, 5, 10000),
672 2 : (1208592, 5, 10000),
673 2 : (1208885, 5, 10000),
674 2 : (1208873, 5, 10000),
675 2 : (1209128, 5, 10000)
676 2 : ]
677 2 : );
678 2 :
679 2 : tmpdir.close().unwrap();
680 2 : }
681 :
682 : #[tokio::test]
683 2 : async fn verify_parquet_unreliable_upload() {
684 2 : let tmpdir = camino_tempfile::tempdir().unwrap();
685 2 :
686 2 : let config = ParquetConfig {
687 2 : propeties: Arc::new(WriterProperties::new()),
688 2 : rows_per_group: 2_000,
689 2 : file_size: 1_000_000,
690 2 : max_duration: time::Duration::from_secs(20 * 60),
691 2 : test_remote_failures: 2,
692 2 : };
693 2 :
694 2 : let rx = random_stream(50_000);
695 124 : let file_stats = run_test(tmpdir.path(), config, rx).await;
696 2 :
697 2 : assert_eq!(
698 2 : file_stats,
699 2 : [
700 2 : (1315874, 3, 6000),
701 2 : (1315867, 3, 6000),
702 2 : (1315927, 3, 6000),
703 2 : (1315884, 3, 6000),
704 2 : (1316014, 3, 6000),
705 2 : (1315856, 3, 6000),
706 2 : (1315648, 3, 6000),
707 2 : (1315884, 3, 6000),
708 2 : (438913, 1, 2000)
709 2 : ]
710 2 : );
711 2 :
712 2 : tmpdir.close().unwrap();
713 2 : }
714 :
715 : #[tokio::test(start_paused = true)]
716 2 : async fn verify_parquet_regular_upload() {
717 2 : let tmpdir = camino_tempfile::tempdir().unwrap();
718 2 :
719 2 : let config = ParquetConfig {
720 2 : propeties: Arc::new(WriterProperties::new()),
721 2 : rows_per_group: 2_000,
722 2 : file_size: 1_000_000,
723 2 : max_duration: time::Duration::from_secs(60),
724 2 : test_remote_failures: 2,
725 2 : };
726 2 :
727 2 : let (tx, mut rx) = mpsc::unbounded_channel();
728 2 :
729 2 : tokio::spawn(async move {
730 8 : for _ in 0..3 {
731 6 : let mut s = random_stream(3000);
732 18006 : while let Some(r) = s.next().await {
733 18000 : tx.send(r).unwrap();
734 18000 : }
735 6 : time::sleep(time::Duration::from_secs(70)).await
736 2 : }
737 2 : });
738 2 :
739 18142 : let rx = futures::stream::poll_fn(move |cx| rx.poll_recv(cx));
740 178 : let file_stats = run_test(tmpdir.path(), config, rx).await;
741 2 :
742 2 : // files are smaller than the size threshold, but they took too long to fill so were flushed early
743 2 : assert_eq!(
744 2 : file_stats,
745 2 : [(659836, 2, 3001), (659550, 2, 3000), (659346, 2, 2999)]
746 2 : );
747 2 :
748 2 : tmpdir.close().unwrap();
749 2 : }
750 : }
|
