Line data Source code
1 : use std::sync::Arc;
2 : use std::time::SystemTime;
3 :
4 : use anyhow::Context;
5 : use bytes::buf::Writer;
6 : use bytes::{BufMut, BytesMut};
7 : use chrono::{Datelike, Timelike};
8 : use futures::{Stream, StreamExt};
9 : use parquet::basic::Compression;
10 : use parquet::file::metadata::RowGroupMetaDataPtr;
11 : use parquet::file::properties::{DEFAULT_PAGE_SIZE, WriterProperties, WriterPropertiesPtr};
12 : use parquet::file::writer::SerializedFileWriter;
13 : use parquet::record::RecordWriter;
14 : use pq_proto::StartupMessageParams;
15 : use remote_storage::{GenericRemoteStorage, RemotePath, RemoteStorageConfig, TimeoutOrCancel};
16 : use serde::ser::SerializeMap;
17 : use tokio::sync::mpsc;
18 : use tokio::time;
19 : use tokio_util::sync::CancellationToken;
20 : use tracing::{Span, debug, info};
21 : use utils::backoff;
22 :
23 : use super::{LOG_CHAN, RequestContextInner};
24 : use crate::config::remote_storage_from_toml;
25 : use crate::context::LOG_CHAN_DISCONNECT;
26 : use crate::ext::TaskExt;
27 :
28 : #[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 1272000 : #[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 : user_agent: Option<String>,
86 : endpoint_id: Option<String>,
87 : database: Option<String>,
88 : project: Option<String>,
89 : branch: Option<String>,
90 : pg_options: Option<String>,
91 : auth_method: Option<&'static str>,
92 : jwt_issuer: Option<String>,
93 :
94 : error: Option<&'static str>,
95 : /// Success is counted if we form a HTTP response with sql rows inside
96 : /// Or if we make it to proxy_pass
97 : success: bool,
98 : /// Indicates if the cplane started the new compute node for this request.
99 : cold_start_info: &'static str,
100 : /// Tracks time from session start (HTTP request/libpq TCP handshake)
101 : /// Through to success/failure
102 : duration_us: u64,
103 : /// If the session was successful after the disconnect, will be created one more event with filled `disconnect_timestamp`.
104 : disconnect_timestamp: Option<chrono::NaiveDateTime>,
105 : }
106 :
107 : struct Options<'a> {
108 : options: &'a StartupMessageParams,
109 : }
110 :
111 : impl serde::Serialize for Options<'_> {
112 0 : fn serialize<S>(&self, s: S) -> Result<S::Ok, S::Error>
113 0 : where
114 0 : S: serde::Serializer,
115 0 : {
116 0 : let mut state = s.serialize_map(None)?;
117 0 : for (k, v) in self.options.iter() {
118 0 : state.serialize_entry(k, v)?;
119 : }
120 0 : state.end()
121 0 : }
122 : }
123 :
124 : impl From<&RequestContextInner> for RequestData {
125 0 : fn from(value: &RequestContextInner) -> Self {
126 0 : Self {
127 0 : session_id: value.session_id,
128 0 : peer_addr: value.conn_info.addr.ip().to_string(),
129 0 : timestamp: value.first_packet.naive_utc(),
130 0 : username: value.user.as_deref().map(String::from),
131 0 : application_name: value.application.as_deref().map(String::from),
132 0 : user_agent: value.user_agent.as_deref().map(String::from),
133 0 : endpoint_id: value.endpoint_id.as_deref().map(String::from),
134 0 : database: value.dbname.as_deref().map(String::from),
135 0 : project: value.project.as_deref().map(String::from),
136 0 : branch: value.branch.as_deref().map(String::from),
137 0 : pg_options: value
138 0 : .pg_options
139 0 : .as_ref()
140 0 : .and_then(|options| serde_json::to_string(&Options { options }).ok()),
141 0 : auth_method: value.auth_method.as_ref().map(|x| match x {
142 0 : super::AuthMethod::ConsoleRedirect => "console_redirect",
143 0 : super::AuthMethod::ScramSha256 => "scram_sha_256",
144 0 : super::AuthMethod::ScramSha256Plus => "scram_sha_256_plus",
145 0 : super::AuthMethod::Cleartext => "cleartext",
146 0 : super::AuthMethod::Jwt => "jwt",
147 0 : }),
148 0 : jwt_issuer: value.jwt_issuer.clone(),
149 0 : protocol: value.protocol.as_str(),
150 0 : region: value.region,
151 0 : error: value.error_kind.as_ref().map(|e| e.to_metric_label()),
152 0 : success: value.success,
153 0 : cold_start_info: value.cold_start_info.as_str(),
154 0 : duration_us: SystemTime::from(value.first_packet)
155 0 : .elapsed()
156 0 : .unwrap_or_default()
157 0 : .as_micros() as u64, // 584 millenia... good enough
158 0 : disconnect_timestamp: value.disconnect_timestamp.map(|x| x.naive_utc()),
159 0 : }
160 0 : }
161 : }
162 :
163 : /// Parquet request context worker
164 : ///
165 : /// It listened on a channel for all completed requests, extracts the data and writes it into a parquet file,
166 : /// then uploads a completed batch to S3
167 0 : pub async fn worker(
168 0 : cancellation_token: CancellationToken,
169 0 : config: ParquetUploadArgs,
170 0 : ) -> anyhow::Result<()> {
171 0 : let Some(remote_storage_config) = config.parquet_upload_remote_storage else {
172 0 : tracing::warn!("parquet request upload: no s3 bucket configured");
173 0 : return Ok(());
174 : };
175 :
176 0 : let (tx, mut rx) = mpsc::unbounded_channel();
177 0 : LOG_CHAN
178 0 : .set(tx.downgrade())
179 0 : .expect("only one worker should set the channel");
180 0 :
181 0 : // setup row stream that will close on cancellation
182 0 : let cancellation_token2 = cancellation_token.clone();
183 0 : tokio::spawn(async move {
184 0 : cancellation_token2.cancelled().await;
185 : // dropping this sender will cause the channel to close only once
186 : // all the remaining inflight requests have been completed.
187 0 : drop(tx);
188 0 : });
189 0 : let rx = futures::stream::poll_fn(move |cx| rx.poll_recv(cx));
190 0 : let rx = rx.map(RequestData::from);
191 :
192 0 : let storage = GenericRemoteStorage::from_config(&remote_storage_config)
193 0 : .await
194 0 : .context("remote storage init")?;
195 :
196 0 : let properties = WriterProperties::builder()
197 0 : .set_data_page_size_limit(config.parquet_upload_page_size)
198 0 : .set_compression(config.parquet_upload_compression);
199 0 :
200 0 : let parquet_config = ParquetConfig {
201 0 : propeties: Arc::new(properties.build()),
202 0 : rows_per_group: config.parquet_upload_row_group_size,
203 0 : file_size: config.parquet_upload_size,
204 0 : max_duration: config.parquet_upload_maximum_duration,
205 0 :
206 0 : #[cfg(any(test, feature = "testing"))]
207 0 : test_remote_failures: 0,
208 0 : };
209 :
210 : // TODO(anna): consider moving this to a separate function.
211 0 : if let Some(disconnect_events_storage_config) =
212 0 : config.parquet_upload_disconnect_events_remote_storage
213 : {
214 0 : let (tx_disconnect, mut rx_disconnect) = mpsc::unbounded_channel();
215 0 : LOG_CHAN_DISCONNECT
216 0 : .set(tx_disconnect.downgrade())
217 0 : .expect("only one worker should set the channel");
218 0 :
219 0 : // setup row stream that will close on cancellation
220 0 : tokio::spawn(async move {
221 0 : cancellation_token.cancelled().await;
222 : // dropping this sender will cause the channel to close only once
223 : // all the remaining inflight requests have been completed.
224 0 : drop(tx_disconnect);
225 0 : });
226 0 : let rx_disconnect = futures::stream::poll_fn(move |cx| rx_disconnect.poll_recv(cx));
227 0 : let rx_disconnect = rx_disconnect.map(RequestData::from);
228 :
229 0 : let storage_disconnect =
230 0 : GenericRemoteStorage::from_config(&disconnect_events_storage_config)
231 0 : .await
232 0 : .context("remote storage for disconnect events init")?;
233 0 : let parquet_config_disconnect = parquet_config.clone();
234 0 : tokio::try_join!(
235 0 : worker_inner(storage, rx, parquet_config),
236 0 : worker_inner(storage_disconnect, rx_disconnect, parquet_config_disconnect)
237 0 : )
238 0 : .map(|_| ())
239 : } else {
240 0 : worker_inner(storage, rx, parquet_config).await
241 : }
242 0 : }
243 :
244 : #[derive(Clone, Debug)]
245 : struct ParquetConfig {
246 : propeties: WriterPropertiesPtr,
247 : rows_per_group: usize,
248 : file_size: i64,
249 :
250 : max_duration: tokio::time::Duration,
251 :
252 : #[cfg(any(test, feature = "testing"))]
253 : test_remote_failures: u64,
254 : }
255 :
256 4 : async fn worker_inner(
257 4 : storage: GenericRemoteStorage,
258 4 : rx: impl Stream<Item = RequestData>,
259 4 : config: ParquetConfig,
260 4 : ) -> anyhow::Result<()> {
261 : #[cfg(any(test, feature = "testing"))]
262 4 : let storage = if config.test_remote_failures > 0 {
263 2 : GenericRemoteStorage::unreliable_wrapper(storage, config.test_remote_failures)
264 : } else {
265 2 : storage
266 : };
267 :
268 4 : let mut rx = std::pin::pin!(rx);
269 4 :
270 4 : let mut rows = Vec::with_capacity(config.rows_per_group);
271 :
272 4 : let schema = rows.as_slice().schema()?;
273 4 : let buffer = BytesMut::new();
274 4 : let w = buffer.writer();
275 4 : let mut w = SerializedFileWriter::new(w, schema.clone(), config.propeties.clone())?;
276 :
277 4 : let mut last_upload = time::Instant::now();
278 4 :
279 4 : let mut len = 0;
280 159004 : while let Some(row) = rx.next().await {
281 159000 : rows.push(row);
282 159000 : let force = last_upload.elapsed() > config.max_duration;
283 159000 : if rows.len() == config.rows_per_group || force {
284 : let rg_meta;
285 80 : (rows, w, rg_meta) = flush_rows(rows, w).await?;
286 80 : len += rg_meta.compressed_size();
287 158920 : }
288 159000 : if len > config.file_size || force {
289 23 : last_upload = time::Instant::now();
290 23 : let file = upload_parquet(w, len, &storage).await?;
291 23 : w = SerializedFileWriter::new(file, schema.clone(), config.propeties.clone())?;
292 23 : len = 0;
293 158977 : }
294 : }
295 :
296 4 : if !rows.is_empty() {
297 : let rg_meta;
298 1 : (_, w, rg_meta) = flush_rows(rows, w).await?;
299 1 : len += rg_meta.compressed_size();
300 3 : }
301 :
302 4 : if !w.flushed_row_groups().is_empty() {
303 3 : let _rtchk: Writer<BytesMut> = upload_parquet(w, len, &storage).await?;
304 1 : }
305 :
306 4 : Ok(())
307 4 : }
308 :
309 81 : async fn flush_rows<W>(
310 81 : rows: Vec<RequestData>,
311 81 : mut w: SerializedFileWriter<W>,
312 81 : ) -> anyhow::Result<(
313 81 : Vec<RequestData>,
314 81 : SerializedFileWriter<W>,
315 81 : RowGroupMetaDataPtr,
316 81 : )>
317 81 : where
318 81 : W: std::io::Write + Send + 'static,
319 81 : {
320 81 : let span = Span::current();
321 81 : let (mut rows, w, rg_meta) = tokio::task::spawn_blocking(move || {
322 81 : let _enter = span.enter();
323 :
324 81 : let mut rg = w.next_row_group()?;
325 81 : rows.as_slice().write_to_row_group(&mut rg)?;
326 81 : let rg_meta = rg.close()?;
327 :
328 81 : let size = rg_meta.compressed_size();
329 81 : let compression = rg_meta.compressed_size() as f64 / rg_meta.total_byte_size() as f64;
330 81 :
331 81 : debug!(size, compression, "flushed row group to parquet file");
332 :
333 81 : Ok::<_, parquet::errors::ParquetError>((rows, w, rg_meta))
334 81 : })
335 81 : .await
336 81 : .propagate_task_panic()?;
337 :
338 81 : rows.clear();
339 81 : Ok((rows, w, rg_meta))
340 81 : }
341 :
342 26 : async fn upload_parquet(
343 26 : mut w: SerializedFileWriter<Writer<BytesMut>>,
344 26 : len: i64,
345 26 : storage: &GenericRemoteStorage,
346 26 : ) -> anyhow::Result<Writer<BytesMut>> {
347 26 : let len_uncompressed = w
348 26 : .flushed_row_groups()
349 26 : .iter()
350 81 : .map(|rg| rg.total_byte_size())
351 26 : .sum::<i64>();
352 :
353 : // I don't know how compute intensive this is, although it probably isn't much... better be safe than sorry.
354 : // finish method only available on the fork: https://github.com/apache/arrow-rs/issues/5253
355 26 : let (mut buffer, metadata) =
356 26 : tokio::task::spawn_blocking(move || -> parquet::errors::Result<_> {
357 26 : let metadata = w.finish()?;
358 26 : let buffer = std::mem::take(w.inner_mut().get_mut());
359 26 : Ok((buffer, metadata))
360 26 : })
361 26 : .await
362 26 : .propagate_task_panic()?;
363 :
364 26 : let data = buffer.split().freeze();
365 26 :
366 26 : let compression = len as f64 / len_uncompressed as f64;
367 26 : let size = data.len();
368 26 : let now = chrono::Utc::now();
369 26 : let id = uuid::Uuid::new_v7(uuid::Timestamp::from_unix(
370 26 : uuid::NoContext,
371 26 : // we won't be running this in 1970. this cast is ok
372 26 : now.timestamp() as u64,
373 26 : now.timestamp_subsec_nanos(),
374 26 : ));
375 26 :
376 26 : info!(
377 : %id,
378 : rows = metadata.num_rows,
379 0 : size, compression, "uploading request parquet file"
380 : );
381 :
382 26 : let year = now.year();
383 26 : let month = now.month();
384 26 : let day = now.day();
385 26 : let hour = now.hour();
386 : // segment files by time for S3 performance
387 26 : let path = RemotePath::from_string(&format!(
388 26 : "{year:04}/{month:02}/{day:02}/{hour:02}/requests_{id}.parquet"
389 26 : ))?;
390 26 : let cancel = CancellationToken::new();
391 26 : let maybe_err = backoff::retry(
392 38 : || async {
393 38 : let stream = futures::stream::once(futures::future::ready(Ok(data.clone())));
394 38 : storage
395 38 : .upload(stream, data.len(), &path, None, &cancel)
396 38 : .await
397 76 : },
398 26 : TimeoutOrCancel::caused_by_cancel,
399 26 : FAILED_UPLOAD_WARN_THRESHOLD,
400 26 : FAILED_UPLOAD_MAX_RETRIES,
401 26 : "request_data_upload",
402 26 : // we don't want cancellation to interrupt here, so we make a dummy cancel token
403 26 : &cancel,
404 26 : )
405 26 : .await
406 26 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
407 26 : .and_then(|x| x)
408 26 : .with_context(|| format!("request_data_upload: path={path}"))
409 26 : .err();
410 :
411 26 : if let Some(err) = maybe_err {
412 0 : tracing::error!(%id, %path, error = ?err, "failed to upload request data");
413 26 : }
414 :
415 26 : Ok(buffer.writer())
416 26 : }
417 :
418 : #[cfg(test)]
419 : #[expect(clippy::unwrap_used)]
420 : mod tests {
421 : use std::net::Ipv4Addr;
422 : use std::num::NonZeroUsize;
423 : use std::sync::Arc;
424 :
425 : use camino::Utf8Path;
426 : use clap::Parser;
427 : use futures::{Stream, StreamExt};
428 : use itertools::Itertools;
429 : use parquet::basic::{Compression, ZstdLevel};
430 : use parquet::file::properties::{DEFAULT_PAGE_SIZE, WriterProperties};
431 : use parquet::file::reader::FileReader;
432 : use parquet::file::serialized_reader::SerializedFileReader;
433 : use rand::rngs::StdRng;
434 : use rand::{Rng, SeedableRng};
435 : use remote_storage::{
436 : DEFAULT_MAX_KEYS_PER_LIST_RESPONSE, DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT,
437 : GenericRemoteStorage, RemoteStorageConfig, RemoteStorageKind, S3Config,
438 : };
439 : use tokio::sync::mpsc;
440 : use tokio::time;
441 : use walkdir::WalkDir;
442 :
443 : use super::{ParquetConfig, ParquetUploadArgs, RequestData, worker_inner};
444 :
445 : #[derive(Parser)]
446 : struct ProxyCliArgs {
447 : #[clap(flatten)]
448 : parquet_upload: ParquetUploadArgs,
449 : }
450 :
451 : #[test]
452 1 : fn default_parser() {
453 1 : let ProxyCliArgs { parquet_upload } = ProxyCliArgs::parse_from(["proxy"]);
454 1 : assert_eq!(parquet_upload.parquet_upload_remote_storage, None);
455 1 : assert_eq!(parquet_upload.parquet_upload_row_group_size, 8192);
456 1 : assert_eq!(parquet_upload.parquet_upload_page_size, DEFAULT_PAGE_SIZE);
457 1 : assert_eq!(parquet_upload.parquet_upload_size, 100_000_000);
458 1 : assert_eq!(
459 1 : parquet_upload.parquet_upload_maximum_duration,
460 1 : time::Duration::from_secs(20 * 60)
461 1 : );
462 1 : assert_eq!(
463 1 : parquet_upload.parquet_upload_compression,
464 1 : Compression::UNCOMPRESSED
465 1 : );
466 1 : }
467 :
468 : #[test]
469 1 : fn full_parser() {
470 1 : let ProxyCliArgs { parquet_upload } = ProxyCliArgs::parse_from([
471 1 : "proxy",
472 1 : "--parquet-upload-remote-storage",
473 1 : "{bucket_name='default',prefix_in_bucket='proxy/',bucket_region='us-east-1',endpoint='http://minio:9000'}",
474 1 : "--parquet-upload-row-group-size",
475 1 : "100",
476 1 : "--parquet-upload-page-size",
477 1 : "10000",
478 1 : "--parquet-upload-size",
479 1 : "10000000",
480 1 : "--parquet-upload-maximum-duration",
481 1 : "10m",
482 1 : "--parquet-upload-compression",
483 1 : "zstd(5)",
484 1 : ]);
485 1 : assert_eq!(
486 1 : parquet_upload.parquet_upload_remote_storage,
487 1 : Some(RemoteStorageConfig {
488 1 : storage: RemoteStorageKind::AwsS3(S3Config {
489 1 : bucket_name: "default".into(),
490 1 : bucket_region: "us-east-1".into(),
491 1 : prefix_in_bucket: Some("proxy/".into()),
492 1 : endpoint: Some("http://minio:9000".into()),
493 1 : concurrency_limit: NonZeroUsize::new(
494 1 : DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT
495 1 : )
496 1 : .unwrap(),
497 1 : max_keys_per_list_response: DEFAULT_MAX_KEYS_PER_LIST_RESPONSE,
498 1 : upload_storage_class: None,
499 1 : }),
500 1 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
501 1 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
502 1 : })
503 1 : );
504 1 : assert_eq!(parquet_upload.parquet_upload_row_group_size, 100);
505 1 : assert_eq!(parquet_upload.parquet_upload_page_size, 10000);
506 1 : assert_eq!(parquet_upload.parquet_upload_size, 10_000_000);
507 1 : assert_eq!(
508 1 : parquet_upload.parquet_upload_maximum_duration,
509 1 : time::Duration::from_secs(10 * 60)
510 1 : );
511 1 : assert_eq!(
512 1 : parquet_upload.parquet_upload_compression,
513 1 : Compression::ZSTD(ZstdLevel::try_new(5).unwrap())
514 1 : );
515 1 : }
516 :
517 159000 : fn generate_request_data(rng: &mut impl Rng) -> RequestData {
518 159000 : RequestData {
519 159000 : session_id: uuid::Builder::from_random_bytes(rng.r#gen()).into_uuid(),
520 159000 : peer_addr: Ipv4Addr::from(rng.r#gen::<[u8; 4]>()).to_string(),
521 159000 : timestamp: chrono::DateTime::from_timestamp_millis(
522 159000 : rng.gen_range(1703862754..1803862754),
523 159000 : )
524 159000 : .unwrap()
525 159000 : .naive_utc(),
526 159000 : application_name: Some("test".to_owned()),
527 159000 : user_agent: Some("test-user-agent".to_owned()),
528 159000 : username: Some(hex::encode(rng.r#gen::<[u8; 4]>())),
529 159000 : endpoint_id: Some(hex::encode(rng.r#gen::<[u8; 16]>())),
530 159000 : database: Some(hex::encode(rng.r#gen::<[u8; 16]>())),
531 159000 : project: Some(hex::encode(rng.r#gen::<[u8; 16]>())),
532 159000 : branch: Some(hex::encode(rng.r#gen::<[u8; 16]>())),
533 159000 : pg_options: None,
534 159000 : auth_method: None,
535 159000 : jwt_issuer: None,
536 159000 : protocol: ["tcp", "ws", "http"][rng.gen_range(0..3)],
537 159000 : region: "us-east-1",
538 159000 : error: None,
539 159000 : success: rng.r#gen(),
540 159000 : cold_start_info: "no",
541 159000 : duration_us: rng.gen_range(0..30_000_000),
542 159000 : disconnect_timestamp: None,
543 159000 : }
544 159000 : }
545 :
546 6 : fn random_stream(len: usize) -> impl Stream<Item = RequestData> + Unpin {
547 6 : let mut rng = StdRng::from_seed([0x39; 32]);
548 6 : futures::stream::iter(
549 159000 : std::iter::repeat_with(move || generate_request_data(&mut rng)).take(len),
550 6 : )
551 6 : }
552 :
553 4 : async fn run_test(
554 4 : tmpdir: &Utf8Path,
555 4 : config: ParquetConfig,
556 4 : rx: impl Stream<Item = RequestData>,
557 4 : ) -> Vec<(u64, usize, i64)> {
558 4 : let remote_storage_config = RemoteStorageConfig {
559 4 : storage: RemoteStorageKind::LocalFs {
560 4 : local_path: tmpdir.to_path_buf(),
561 4 : },
562 4 : timeout: std::time::Duration::from_secs(120),
563 4 : small_timeout: std::time::Duration::from_secs(30),
564 4 : };
565 4 : let storage = GenericRemoteStorage::from_config(&remote_storage_config)
566 4 : .await
567 4 : .unwrap();
568 4 :
569 4 : worker_inner(storage, rx, config).await.unwrap();
570 4 :
571 4 : let mut files = WalkDir::new(tmpdir.as_std_path())
572 4 : .into_iter()
573 46 : .filter_map(|entry| entry.ok())
574 46 : .filter(|entry| entry.file_type().is_file())
575 26 : .map(|entry| entry.path().to_path_buf())
576 4 : .collect_vec();
577 4 : files.sort();
578 4 :
579 4 : files
580 4 : .into_iter()
581 26 : .map(|path| std::fs::File::open(tmpdir.as_std_path().join(path)).unwrap())
582 26 : .map(|file| {
583 26 : (
584 26 : file.metadata().unwrap(),
585 26 : SerializedFileReader::new(file).unwrap().metadata().clone(),
586 26 : )
587 26 : })
588 26 : .map(|(file_meta, parquet_meta)| {
589 26 : (
590 26 : file_meta.len(),
591 26 : parquet_meta.num_row_groups(),
592 26 : parquet_meta.file_metadata().num_rows(),
593 26 : )
594 26 : })
595 4 : .collect()
596 4 : }
597 :
598 : #[tokio::test]
599 1 : async fn verify_parquet_no_compression() {
600 1 : let tmpdir = camino_tempfile::tempdir().unwrap();
601 1 :
602 1 : let config = ParquetConfig {
603 1 : propeties: Arc::new(WriterProperties::new()),
604 1 : rows_per_group: 2_000,
605 1 : file_size: 1_000_000,
606 1 : max_duration: time::Duration::from_secs(20 * 60),
607 1 : test_remote_failures: 0,
608 1 : };
609 1 :
610 1 : let rx = random_stream(50_000);
611 1 : let file_stats = run_test(tmpdir.path(), config, rx).await;
612 1 :
613 1 : assert_eq!(
614 1 : file_stats,
615 1 : [
616 1 : (1313953, 3, 6000),
617 1 : (1313942, 3, 6000),
618 1 : (1314001, 3, 6000),
619 1 : (1313958, 3, 6000),
620 1 : (1314094, 3, 6000),
621 1 : (1313931, 3, 6000),
622 1 : (1313725, 3, 6000),
623 1 : (1313960, 3, 6000),
624 1 : (438318, 1, 2000)
625 1 : ]
626 1 : );
627 1 :
628 1 : tmpdir.close().unwrap();
629 1 : }
630 :
631 : #[tokio::test]
632 1 : async fn verify_parquet_strong_compression() {
633 1 : let tmpdir = camino_tempfile::tempdir().unwrap();
634 1 :
635 1 : let config = ParquetConfig {
636 1 : propeties: Arc::new(
637 1 : WriterProperties::builder()
638 1 : .set_compression(parquet::basic::Compression::ZSTD(
639 1 : ZstdLevel::try_new(10).unwrap(),
640 1 : ))
641 1 : .build(),
642 1 : ),
643 1 : rows_per_group: 2_000,
644 1 : file_size: 1_000_000,
645 1 : max_duration: time::Duration::from_secs(20 * 60),
646 1 : test_remote_failures: 0,
647 1 : };
648 1 :
649 1 : let rx = random_stream(50_000);
650 1 : let file_stats = run_test(tmpdir.path(), config, rx).await;
651 1 :
652 1 : // with strong compression, the files are smaller
653 1 : assert_eq!(
654 1 : file_stats,
655 1 : [
656 1 : (1205810, 5, 10000),
657 1 : (1205534, 5, 10000),
658 1 : (1205835, 5, 10000),
659 1 : (1205820, 5, 10000),
660 1 : (1206074, 5, 10000)
661 1 : ]
662 1 : );
663 1 :
664 1 : tmpdir.close().unwrap();
665 1 : }
666 :
667 : #[tokio::test]
668 1 : async fn verify_parquet_unreliable_upload() {
669 1 : let tmpdir = camino_tempfile::tempdir().unwrap();
670 1 :
671 1 : let config = ParquetConfig {
672 1 : propeties: Arc::new(WriterProperties::new()),
673 1 : rows_per_group: 2_000,
674 1 : file_size: 1_000_000,
675 1 : max_duration: time::Duration::from_secs(20 * 60),
676 1 : test_remote_failures: 2,
677 1 : };
678 1 :
679 1 : let rx = random_stream(50_000);
680 1 : let file_stats = run_test(tmpdir.path(), config, rx).await;
681 1 :
682 1 : assert_eq!(
683 1 : file_stats,
684 1 : [
685 1 : (1313953, 3, 6000),
686 1 : (1313942, 3, 6000),
687 1 : (1314001, 3, 6000),
688 1 : (1313958, 3, 6000),
689 1 : (1314094, 3, 6000),
690 1 : (1313931, 3, 6000),
691 1 : (1313725, 3, 6000),
692 1 : (1313960, 3, 6000),
693 1 : (438318, 1, 2000)
694 1 : ]
695 1 : );
696 1 :
697 1 : tmpdir.close().unwrap();
698 1 : }
699 :
700 : #[tokio::test(start_paused = true)]
701 1 : async fn verify_parquet_regular_upload() {
702 1 : let tmpdir = camino_tempfile::tempdir().unwrap();
703 1 :
704 1 : let config = ParquetConfig {
705 1 : propeties: Arc::new(WriterProperties::new()),
706 1 : rows_per_group: 2_000,
707 1 : file_size: 1_000_000,
708 1 : max_duration: time::Duration::from_secs(60),
709 1 : test_remote_failures: 2,
710 1 : };
711 1 :
712 1 : let (tx, mut rx) = mpsc::unbounded_channel();
713 1 :
714 1 : tokio::spawn(async move {
715 4 : for _ in 0..3 {
716 3 : let mut s = random_stream(3000);
717 9003 : while let Some(r) = s.next().await {
718 9000 : tx.send(r).unwrap();
719 9000 : }
720 3 : time::sleep(time::Duration::from_secs(70)).await;
721 1 : }
722 1 : });
723 1 :
724 9071 : let rx = futures::stream::poll_fn(move |cx| rx.poll_recv(cx));
725 1 : let file_stats = run_test(tmpdir.path(), config, rx).await;
726 1 :
727 1 : // files are smaller than the size threshold, but they took too long to fill so were flushed early
728 1 : assert_eq!(
729 1 : file_stats,
730 1 : [(658584, 2, 3001), (658298, 2, 3000), (658094, 2, 2999)]
731 1 : );
732 1 :
733 1 : tmpdir.close().unwrap();
734 1 : }
735 : }
|
