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::{RequestMonitoringInner, LOG_CHAN};
27 :
28 6 : #[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 3 : #[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 3 : #[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 3 : #[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 3 : #[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(crate) const FAILED_UPLOAD_WARN_THRESHOLD: u32 = 3;
66 : pub(crate) 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 3180085 : #[derive(parquet_derive::ParquetRecordWriter)]
76 : pub(crate) 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<&RequestMonitoringInner> for RequestData {
122 0 : fn from(value: &RequestMonitoringInner) -> 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 = GenericRemoteStorage::from_config(&remote_storage_config)
185 0 : .await
186 0 : .context("remote storage init")?;
187 :
188 0 : let properties = WriterProperties::builder()
189 0 : .set_data_page_size_limit(config.parquet_upload_page_size)
190 0 : .set_compression(config.parquet_upload_compression);
191 0 :
192 0 : let parquet_config = ParquetConfig {
193 0 : propeties: Arc::new(properties.build()),
194 0 : rows_per_group: config.parquet_upload_row_group_size,
195 0 : file_size: config.parquet_upload_size,
196 0 : max_duration: config.parquet_upload_maximum_duration,
197 0 :
198 0 : #[cfg(any(test, feature = "testing"))]
199 0 : test_remote_failures: 0,
200 0 : };
201 :
202 : // TODO(anna): consider moving this to a separate function.
203 0 : if let Some(disconnect_events_storage_config) =
204 0 : config.parquet_upload_disconnect_events_remote_storage
205 : {
206 0 : let (tx_disconnect, mut rx_disconnect) = mpsc::unbounded_channel();
207 0 : LOG_CHAN_DISCONNECT.set(tx_disconnect.downgrade()).unwrap();
208 0 :
209 0 : // setup row stream that will close on cancellation
210 0 : tokio::spawn(async move {
211 0 : cancellation_token.cancelled().await;
212 : // dropping this sender will cause the channel to close only once
213 : // all the remaining inflight requests have been completed.
214 0 : drop(tx_disconnect);
215 0 : });
216 0 : let rx_disconnect = futures::stream::poll_fn(move |cx| rx_disconnect.poll_recv(cx));
217 0 : let rx_disconnect = rx_disconnect.map(RequestData::from);
218 :
219 0 : let storage_disconnect =
220 0 : GenericRemoteStorage::from_config(&disconnect_events_storage_config)
221 0 : .await
222 0 : .context("remote storage for disconnect events init")?;
223 0 : let parquet_config_disconnect = parquet_config.clone();
224 0 : tokio::try_join!(
225 0 : worker_inner(storage, rx, parquet_config),
226 0 : worker_inner(storage_disconnect, rx_disconnect, parquet_config_disconnect)
227 0 : )
228 0 : .map(|_| ())
229 : } else {
230 0 : worker_inner(storage, rx, parquet_config).await
231 : }
232 0 : }
233 :
234 : #[derive(Clone, Debug)]
235 : struct ParquetConfig {
236 : propeties: WriterPropertiesPtr,
237 : rows_per_group: usize,
238 : file_size: i64,
239 :
240 : max_duration: tokio::time::Duration,
241 :
242 : #[cfg(any(test, feature = "testing"))]
243 : test_remote_failures: u64,
244 : }
245 :
246 4 : async fn worker_inner(
247 4 : storage: GenericRemoteStorage,
248 4 : rx: impl Stream<Item = RequestData>,
249 4 : config: ParquetConfig,
250 4 : ) -> anyhow::Result<()> {
251 : #[cfg(any(test, feature = "testing"))]
252 4 : let storage = if config.test_remote_failures > 0 {
253 2 : GenericRemoteStorage::unreliable_wrapper(storage, config.test_remote_failures)
254 : } else {
255 2 : storage
256 : };
257 :
258 4 : let mut rx = std::pin::pin!(rx);
259 4 :
260 4 : let mut rows = Vec::with_capacity(config.rows_per_group);
261 :
262 4 : let schema = rows.as_slice().schema()?;
263 4 : let buffer = BytesMut::new();
264 4 : let w = buffer.writer();
265 4 : let mut w = SerializedFileWriter::new(w, schema.clone(), config.propeties.clone())?;
266 :
267 4 : let mut last_upload = time::Instant::now();
268 4 :
269 4 : let mut len = 0;
270 159004 : while let Some(row) = rx.next().await {
271 159000 : rows.push(row);
272 159000 : let force = last_upload.elapsed() > config.max_duration;
273 159000 : if rows.len() == config.rows_per_group || force {
274 : let rg_meta;
275 80 : (rows, w, rg_meta) = flush_rows(rows, w).await?;
276 80 : len += rg_meta.compressed_size();
277 158920 : }
278 159000 : if len > config.file_size || force {
279 23 : last_upload = time::Instant::now();
280 96 : let file = upload_parquet(w, len, &storage).await?;
281 23 : w = SerializedFileWriter::new(file, schema.clone(), config.propeties.clone())?;
282 23 : len = 0;
283 158977 : }
284 : }
285 :
286 4 : if !rows.is_empty() {
287 : let rg_meta;
288 1 : (_, w, rg_meta) = flush_rows(rows, w).await?;
289 1 : len += rg_meta.compressed_size();
290 3 : }
291 :
292 4 : if !w.flushed_row_groups().is_empty() {
293 12 : let _rtchk: Writer<BytesMut> = upload_parquet(w, len, &storage).await?;
294 1 : }
295 :
296 4 : Ok(())
297 4 : }
298 :
299 81 : async fn flush_rows<W>(
300 81 : rows: Vec<RequestData>,
301 81 : mut w: SerializedFileWriter<W>,
302 81 : ) -> anyhow::Result<(
303 81 : Vec<RequestData>,
304 81 : SerializedFileWriter<W>,
305 81 : RowGroupMetaDataPtr,
306 81 : )>
307 81 : where
308 81 : W: std::io::Write + Send + 'static,
309 81 : {
310 81 : let span = Span::current();
311 81 : let (mut rows, w, rg_meta) = tokio::task::spawn_blocking(move || {
312 81 : let _enter = span.enter();
313 :
314 81 : let mut rg = w.next_row_group()?;
315 81 : rows.as_slice().write_to_row_group(&mut rg)?;
316 81 : let rg_meta = rg.close()?;
317 :
318 81 : let size = rg_meta.compressed_size();
319 81 : let compression = rg_meta.compressed_size() as f64 / rg_meta.total_byte_size() as f64;
320 81 :
321 81 : debug!(size, compression, "flushed row group to parquet file");
322 :
323 81 : Ok::<_, parquet::errors::ParquetError>((rows, w, rg_meta))
324 81 : })
325 81 : .await
326 81 : .unwrap()?;
327 :
328 81 : rows.clear();
329 81 : Ok((rows, w, rg_meta))
330 81 : }
331 :
332 26 : async fn upload_parquet(
333 26 : mut w: SerializedFileWriter<Writer<BytesMut>>,
334 26 : len: i64,
335 26 : storage: &GenericRemoteStorage,
336 26 : ) -> anyhow::Result<Writer<BytesMut>> {
337 26 : let len_uncompressed = w
338 26 : .flushed_row_groups()
339 26 : .iter()
340 81 : .map(|rg| rg.total_byte_size())
341 26 : .sum::<i64>();
342 :
343 : // I don't know how compute intensive this is, although it probably isn't much... better be safe than sorry.
344 : // finish method only available on the fork: https://github.com/apache/arrow-rs/issues/5253
345 26 : let (mut buffer, metadata) =
346 26 : tokio::task::spawn_blocking(move || -> parquet::errors::Result<_> {
347 26 : let metadata = w.finish()?;
348 26 : let buffer = std::mem::take(w.inner_mut().get_mut());
349 26 : Ok((buffer, metadata))
350 26 : })
351 26 : .await
352 26 : .unwrap()?;
353 :
354 26 : let data = buffer.split().freeze();
355 26 :
356 26 : let compression = len as f64 / len_uncompressed as f64;
357 26 : let size = data.len();
358 26 : let now = chrono::Utc::now();
359 26 : let id = uuid::Uuid::new_v7(uuid::Timestamp::from_unix(
360 26 : uuid::NoContext,
361 26 : // we won't be running this in 1970. this cast is ok
362 26 : now.timestamp() as u64,
363 26 : now.timestamp_subsec_nanos(),
364 26 : ));
365 26 :
366 26 : info!(
367 : %id,
368 : rows = metadata.num_rows,
369 0 : size, compression, "uploading request parquet file"
370 : );
371 :
372 26 : let year = now.year();
373 26 : let month = now.month();
374 26 : let day = now.day();
375 26 : let hour = now.hour();
376 : // segment files by time for S3 performance
377 26 : let path = RemotePath::from_string(&format!(
378 26 : "{year:04}/{month:02}/{day:02}/{hour:02}/requests_{id}.parquet"
379 26 : ))?;
380 26 : let cancel = CancellationToken::new();
381 26 : let maybe_err = backoff::retry(
382 38 : || async {
383 38 : let stream = futures::stream::once(futures::future::ready(Ok(data.clone())));
384 38 : storage
385 38 : .upload(stream, data.len(), &path, None, &cancel)
386 82 : .await
387 76 : },
388 26 : TimeoutOrCancel::caused_by_cancel,
389 26 : FAILED_UPLOAD_WARN_THRESHOLD,
390 26 : FAILED_UPLOAD_MAX_RETRIES,
391 26 : "request_data_upload",
392 26 : // we don't want cancellation to interrupt here, so we make a dummy cancel token
393 26 : &cancel,
394 26 : )
395 82 : .await
396 26 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
397 26 : .and_then(|x| x)
398 26 : .context("request_data_upload")
399 26 : .err();
400 :
401 26 : if let Some(err) = maybe_err {
402 0 : tracing::warn!(%id, %err, "failed to upload request data");
403 26 : }
404 :
405 26 : Ok(buffer.writer())
406 26 : }
407 :
408 : #[cfg(test)]
409 : mod tests {
410 : use std::{net::Ipv4Addr, num::NonZeroUsize, sync::Arc};
411 :
412 : use camino::Utf8Path;
413 : use clap::Parser;
414 : use futures::{Stream, StreamExt};
415 : use itertools::Itertools;
416 : use parquet::{
417 : basic::{Compression, ZstdLevel},
418 : file::{
419 : properties::{WriterProperties, DEFAULT_PAGE_SIZE},
420 : reader::FileReader,
421 : serialized_reader::SerializedFileReader,
422 : },
423 : };
424 : use rand::{rngs::StdRng, Rng, SeedableRng};
425 : use remote_storage::{
426 : GenericRemoteStorage, RemoteStorageConfig, RemoteStorageKind, S3Config,
427 : DEFAULT_MAX_KEYS_PER_LIST_RESPONSE, DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT,
428 : };
429 : use tokio::{sync::mpsc, time};
430 : use walkdir::WalkDir;
431 :
432 : use super::{worker_inner, ParquetConfig, ParquetUploadArgs, RequestData};
433 :
434 2 : #[derive(Parser)]
435 : struct ProxyCliArgs {
436 : #[clap(flatten)]
437 : parquet_upload: ParquetUploadArgs,
438 : }
439 :
440 : #[test]
441 1 : fn default_parser() {
442 1 : let ProxyCliArgs { parquet_upload } = ProxyCliArgs::parse_from(["proxy"]);
443 1 : assert_eq!(parquet_upload.parquet_upload_remote_storage, None);
444 1 : assert_eq!(parquet_upload.parquet_upload_row_group_size, 8192);
445 1 : assert_eq!(parquet_upload.parquet_upload_page_size, DEFAULT_PAGE_SIZE);
446 1 : assert_eq!(parquet_upload.parquet_upload_size, 100_000_000);
447 1 : assert_eq!(
448 1 : parquet_upload.parquet_upload_maximum_duration,
449 1 : time::Duration::from_secs(20 * 60)
450 1 : );
451 1 : assert_eq!(
452 1 : parquet_upload.parquet_upload_compression,
453 1 : Compression::UNCOMPRESSED
454 1 : );
455 1 : }
456 :
457 : #[test]
458 1 : fn full_parser() {
459 1 : let ProxyCliArgs { parquet_upload } = ProxyCliArgs::parse_from([
460 1 : "proxy",
461 1 : "--parquet-upload-remote-storage",
462 1 : "{bucket_name='default',prefix_in_bucket='proxy/',bucket_region='us-east-1',endpoint='http://minio:9000'}",
463 1 : "--parquet-upload-row-group-size",
464 1 : "100",
465 1 : "--parquet-upload-page-size",
466 1 : "10000",
467 1 : "--parquet-upload-size",
468 1 : "10000000",
469 1 : "--parquet-upload-maximum-duration",
470 1 : "10m",
471 1 : "--parquet-upload-compression",
472 1 : "zstd(5)",
473 1 : ]);
474 1 : assert_eq!(
475 1 : parquet_upload.parquet_upload_remote_storage,
476 1 : Some(RemoteStorageConfig {
477 1 : storage: RemoteStorageKind::AwsS3(S3Config {
478 1 : bucket_name: "default".into(),
479 1 : bucket_region: "us-east-1".into(),
480 1 : prefix_in_bucket: Some("proxy/".into()),
481 1 : endpoint: Some("http://minio:9000".into()),
482 1 : concurrency_limit: NonZeroUsize::new(
483 1 : DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT
484 1 : )
485 1 : .unwrap(),
486 1 : max_keys_per_list_response: DEFAULT_MAX_KEYS_PER_LIST_RESPONSE,
487 1 : upload_storage_class: None,
488 1 : }),
489 1 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
490 1 : })
491 1 : );
492 1 : assert_eq!(parquet_upload.parquet_upload_row_group_size, 100);
493 1 : assert_eq!(parquet_upload.parquet_upload_page_size, 10000);
494 1 : assert_eq!(parquet_upload.parquet_upload_size, 10_000_000);
495 1 : assert_eq!(
496 1 : parquet_upload.parquet_upload_maximum_duration,
497 1 : time::Duration::from_secs(10 * 60)
498 1 : );
499 1 : assert_eq!(
500 1 : parquet_upload.parquet_upload_compression,
501 1 : Compression::ZSTD(ZstdLevel::try_new(5).unwrap())
502 1 : );
503 1 : }
504 :
505 159000 : fn generate_request_data(rng: &mut impl Rng) -> RequestData {
506 159000 : RequestData {
507 159000 : session_id: uuid::Builder::from_random_bytes(rng.gen()).into_uuid(),
508 159000 : peer_addr: Ipv4Addr::from(rng.gen::<[u8; 4]>()).to_string(),
509 159000 : timestamp: chrono::DateTime::from_timestamp_millis(
510 159000 : rng.gen_range(1703862754..1803862754),
511 159000 : )
512 159000 : .unwrap()
513 159000 : .naive_utc(),
514 159000 : application_name: Some("test".to_owned()),
515 159000 : username: Some(hex::encode(rng.gen::<[u8; 4]>())),
516 159000 : endpoint_id: Some(hex::encode(rng.gen::<[u8; 16]>())),
517 159000 : database: Some(hex::encode(rng.gen::<[u8; 16]>())),
518 159000 : project: Some(hex::encode(rng.gen::<[u8; 16]>())),
519 159000 : branch: Some(hex::encode(rng.gen::<[u8; 16]>())),
520 159000 : pg_options: None,
521 159000 : auth_method: None,
522 159000 : protocol: ["tcp", "ws", "http"][rng.gen_range(0..3)],
523 159000 : region: "us-east-1",
524 159000 : error: None,
525 159000 : success: rng.gen(),
526 159000 : cold_start_info: "no",
527 159000 : duration_us: rng.gen_range(0..30_000_000),
528 159000 : disconnect_timestamp: None,
529 159000 : }
530 159000 : }
531 :
532 6 : fn random_stream(len: usize) -> impl Stream<Item = RequestData> + Unpin {
533 6 : let mut rng = StdRng::from_seed([0x39; 32]);
534 6 : futures::stream::iter(
535 159000 : std::iter::repeat_with(move || generate_request_data(&mut rng)).take(len),
536 6 : )
537 6 : }
538 :
539 4 : async fn run_test(
540 4 : tmpdir: &Utf8Path,
541 4 : config: ParquetConfig,
542 4 : rx: impl Stream<Item = RequestData>,
543 4 : ) -> Vec<(u64, usize, i64)> {
544 4 : let remote_storage_config = RemoteStorageConfig {
545 4 : storage: RemoteStorageKind::LocalFs {
546 4 : local_path: tmpdir.to_path_buf(),
547 4 : },
548 4 : timeout: std::time::Duration::from_secs(120),
549 4 : };
550 4 : let storage = GenericRemoteStorage::from_config(&remote_storage_config)
551 0 : .await
552 4 : .unwrap();
553 4 :
554 259 : worker_inner(storage, rx, config).await.unwrap();
555 4 :
556 4 : let mut files = WalkDir::new(tmpdir.as_std_path())
557 4 : .into_iter()
558 46 : .filter_map(|entry| entry.ok())
559 46 : .filter(|entry| entry.file_type().is_file())
560 26 : .map(|entry| entry.path().to_path_buf())
561 4 : .collect_vec();
562 4 : files.sort();
563 4 :
564 4 : files
565 4 : .into_iter()
566 26 : .map(|path| std::fs::File::open(tmpdir.as_std_path().join(path)).unwrap())
567 26 : .map(|file| {
568 26 : (
569 26 : file.metadata().unwrap(),
570 26 : SerializedFileReader::new(file).unwrap().metadata().clone(),
571 26 : )
572 26 : })
573 26 : .map(|(file_meta, parquet_meta)| {
574 26 : (
575 26 : file_meta.len(),
576 26 : parquet_meta.num_row_groups(),
577 26 : parquet_meta.file_metadata().num_rows(),
578 26 : )
579 26 : })
580 4 : .collect()
581 4 : }
582 :
583 : #[tokio::test]
584 1 : async fn verify_parquet_no_compression() {
585 1 : let tmpdir = camino_tempfile::tempdir().unwrap();
586 1 :
587 1 : let config = ParquetConfig {
588 1 : propeties: Arc::new(WriterProperties::new()),
589 1 : rows_per_group: 2_000,
590 1 : file_size: 1_000_000,
591 1 : max_duration: time::Duration::from_secs(20 * 60),
592 1 : test_remote_failures: 0,
593 1 : };
594 1 :
595 1 : let rx = random_stream(50_000);
596 62 : let file_stats = run_test(tmpdir.path(), config, rx).await;
597 1 :
598 1 : assert_eq!(
599 1 : file_stats,
600 1 : [
601 1 : (1312632, 3, 6000),
602 1 : (1312621, 3, 6000),
603 1 : (1312680, 3, 6000),
604 1 : (1312637, 3, 6000),
605 1 : (1312773, 3, 6000),
606 1 : (1312610, 3, 6000),
607 1 : (1312404, 3, 6000),
608 1 : (1312639, 3, 6000),
609 1 : (437848, 1, 2000)
610 1 : ]
611 1 : );
612 1 :
613 1 : tmpdir.close().unwrap();
614 1 : }
615 :
616 : #[tokio::test]
617 1 : async fn verify_parquet_strong_compression() {
618 1 : let tmpdir = camino_tempfile::tempdir().unwrap();
619 1 :
620 1 : let config = ParquetConfig {
621 1 : propeties: Arc::new(
622 1 : WriterProperties::builder()
623 1 : .set_compression(parquet::basic::Compression::ZSTD(
624 1 : ZstdLevel::try_new(10).unwrap(),
625 1 : ))
626 1 : .build(),
627 1 : ),
628 1 : rows_per_group: 2_000,
629 1 : file_size: 1_000_000,
630 1 : max_duration: time::Duration::from_secs(20 * 60),
631 1 : test_remote_failures: 0,
632 1 : };
633 1 :
634 1 : let rx = random_stream(50_000);
635 46 : let file_stats = run_test(tmpdir.path(), config, rx).await;
636 1 :
637 1 : // with strong compression, the files are smaller
638 1 : assert_eq!(
639 1 : file_stats,
640 1 : [
641 1 : (1203465, 5, 10000),
642 1 : (1203189, 5, 10000),
643 1 : (1203490, 5, 10000),
644 1 : (1203475, 5, 10000),
645 1 : (1203729, 5, 10000)
646 1 : ]
647 1 : );
648 1 :
649 1 : tmpdir.close().unwrap();
650 1 : }
651 :
652 : #[tokio::test]
653 1 : async fn verify_parquet_unreliable_upload() {
654 1 : let tmpdir = camino_tempfile::tempdir().unwrap();
655 1 :
656 1 : let config = ParquetConfig {
657 1 : propeties: Arc::new(WriterProperties::new()),
658 1 : rows_per_group: 2_000,
659 1 : file_size: 1_000_000,
660 1 : max_duration: time::Duration::from_secs(20 * 60),
661 1 : test_remote_failures: 2,
662 1 : };
663 1 :
664 1 : let rx = random_stream(50_000);
665 62 : let file_stats = run_test(tmpdir.path(), config, rx).await;
666 1 :
667 1 : assert_eq!(
668 1 : file_stats,
669 1 : [
670 1 : (1312632, 3, 6000),
671 1 : (1312621, 3, 6000),
672 1 : (1312680, 3, 6000),
673 1 : (1312637, 3, 6000),
674 1 : (1312773, 3, 6000),
675 1 : (1312610, 3, 6000),
676 1 : (1312404, 3, 6000),
677 1 : (1312639, 3, 6000),
678 1 : (437848, 1, 2000)
679 1 : ]
680 1 : );
681 1 :
682 1 : tmpdir.close().unwrap();
683 1 : }
684 :
685 : #[tokio::test(start_paused = true)]
686 1 : async fn verify_parquet_regular_upload() {
687 1 : let tmpdir = camino_tempfile::tempdir().unwrap();
688 1 :
689 1 : let config = ParquetConfig {
690 1 : propeties: Arc::new(WriterProperties::new()),
691 1 : rows_per_group: 2_000,
692 1 : file_size: 1_000_000,
693 1 : max_duration: time::Duration::from_secs(60),
694 1 : test_remote_failures: 2,
695 1 : };
696 1 :
697 1 : let (tx, mut rx) = mpsc::unbounded_channel();
698 1 :
699 1 : tokio::spawn(async move {
700 4 : for _ in 0..3 {
701 3 : let mut s = random_stream(3000);
702 9003 : while let Some(r) = s.next().await {
703 9000 : tx.send(r).unwrap();
704 9000 : }
705 3 : time::sleep(time::Duration::from_secs(70)).await;
706 1 : }
707 1 : });
708 1 :
709 9071 : let rx = futures::stream::poll_fn(move |cx| rx.poll_recv(cx));
710 89 : let file_stats = run_test(tmpdir.path(), config, rx).await;
711 1 :
712 1 : // files are smaller than the size threshold, but they took too long to fill so were flushed early
713 1 : assert_eq!(
714 1 : file_stats,
715 1 : [(657696, 2, 3001), (657410, 2, 3000), (657206, 2, 2999)]
716 1 : );
717 1 :
718 1 : tmpdir.close().unwrap();
719 1 : }
720 : }
|
