Merge branch 'main' of https://github.com/neondatabase/neon into skyzh/cli-parse-reject

This commit is contained in:
Alex Chi
2023-06-01 13:24:15 -04:00
36 changed files with 1124 additions and 515 deletions

View File

@@ -19,7 +19,29 @@ use crate::reltag::RelTag;
use anyhow::bail;
use bytes::{BufMut, Bytes, BytesMut};
/// A state of a tenant in pageserver's memory.
/// The state of a tenant in this pageserver.
///
/// ```mermaid
/// stateDiagram-v2
///
/// [*] --> Loading: spawn_load()
/// [*] --> Attaching: spawn_attach()
///
/// Loading --> Activating: activate()
/// Attaching --> Activating: activate()
/// Activating --> Active: infallible
///
/// Loading --> Broken: load() failure
/// Attaching --> Broken: attach() failure
///
/// Active --> Stopping: set_stopping(), part of shutdown & detach
/// Stopping --> Broken: late error in remove_tenant_from_memory
///
/// Broken --> [*]: ignore / detach / shutdown
/// Stopping --> [*]: remove_from_memory complete
///
/// Active --> Broken: cfg(testing)-only tenant break point
/// ```
#[derive(
Clone,
PartialEq,
@@ -27,40 +49,63 @@ use bytes::{BufMut, Bytes, BytesMut};
serde::Serialize,
serde::Deserialize,
strum_macros::Display,
strum_macros::EnumString,
strum_macros::EnumVariantNames,
strum_macros::AsRefStr,
strum_macros::IntoStaticStr,
)]
#[serde(tag = "slug", content = "data")]
pub enum TenantState {
/// This tenant is being loaded from local disk
/// This tenant is being loaded from local disk.
///
/// `set_stopping()` and `set_broken()` do not work in this state and wait for it to pass.
Loading,
/// This tenant is being downloaded from cloud storage.
/// This tenant is being attached to the pageserver.
///
/// `set_stopping()` and `set_broken()` do not work in this state and wait for it to pass.
Attaching,
/// Tenant is fully operational
/// The tenant is transitioning from Loading/Attaching to Active.
///
/// While in this state, the individual timelines are being activated.
///
/// `set_stopping()` and `set_broken()` do not work in this state and wait for it to pass.
Activating(ActivatingFrom),
/// The tenant has finished activating and is open for business.
///
/// Transitions out of this state are possible through `set_stopping()` and `set_broken()`.
Active,
/// A tenant is recognized by pageserver, but it is being detached or the
/// The tenant is recognized by pageserver, but it is being detached or the
/// system is being shut down.
///
/// Transitions out of this state are possible through `set_broken()`.
Stopping,
/// A tenant is recognized by the pageserver, but can no longer be used for
/// any operations, because it failed to be activated.
/// The tenant is recognized by the pageserver, but can no longer be used for
/// any operations.
///
/// If the tenant fails to load or attach, it will transition to this state
/// and it is guaranteed that no background tasks are running in its name.
///
/// The other way to transition into this state is from `Stopping` state
/// through `set_broken()` called from `remove_tenant_from_memory()`. That happens
/// if the cleanup future executed by `remove_tenant_from_memory()` fails.
Broken { reason: String, backtrace: String },
}
impl TenantState {
pub fn attachment_status(&self) -> TenantAttachmentStatus {
use TenantAttachmentStatus::*;
// Below TenantState::Activating is used as "transient" or "transparent" state for
// attachment_status determining.
match self {
// The attach procedure writes the marker file before adding the Attaching tenant to the tenants map.
// So, technically, we can return Attached here.
// However, as soon as Console observes Attached, it will proceed with the Postgres-level health check.
// But, our attach task might still be fetching the remote timelines, etc.
// So, return `Maybe` while Attaching, making Console wait for the attach task to finish.
Self::Attaching => Maybe,
Self::Attaching | Self::Activating(ActivatingFrom::Attaching) => Maybe,
// tenant mgr startup distinguishes attaching from loading via marker file.
// If it's loading, there is no attach marker file, i.e., attach had finished in the past.
Self::Loading => Attached,
Self::Loading | Self::Activating(ActivatingFrom::Loading) => Attached,
// We only reach Active after successful load / attach.
// So, call atttachment status Attached.
Self::Active => Attached,
@@ -99,6 +144,15 @@ impl std::fmt::Debug for TenantState {
}
}
/// The only [`TenantState`] variants we could be `TenantState::Activating` from.
#[derive(Clone, Copy, Debug, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
pub enum ActivatingFrom {
/// Arrived to [`TenantState::Activating`] from [`TenantState::Loading`]
Loading,
/// Arrived to [`TenantState::Activating`] from [`TenantState::Attaching`]
Attaching,
}
/// A state of a timeline in pageserver's memory.
#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
pub enum TimelineState {
@@ -187,6 +241,7 @@ pub struct TenantConfig {
pub min_resident_size_override: Option<u64>,
#[clap(long)]
pub evictions_low_residence_duration_metric_threshold: Option<String>,
pub gc_feedback: Option<bool>,
}
impl TenantConfig {
@@ -264,6 +319,7 @@ impl TenantConfigRequest {
eviction_policy: None,
min_resident_size_override: None,
evictions_low_residence_duration_metric_threshold: None,
gc_feedback: None,
};
TenantConfigRequest { tenant_id, config }
}
@@ -871,4 +927,55 @@ mod tests {
err
);
}
#[test]
fn tenantstatus_activating_serde() {
let states = [
TenantState::Activating(ActivatingFrom::Loading),
TenantState::Activating(ActivatingFrom::Attaching),
];
let expected = "[{\"slug\":\"Activating\",\"data\":\"Loading\"},{\"slug\":\"Activating\",\"data\":\"Attaching\"}]";
let actual = serde_json::to_string(&states).unwrap();
assert_eq!(actual, expected);
let parsed = serde_json::from_str::<Vec<TenantState>>(&actual).unwrap();
assert_eq!(states.as_slice(), &parsed);
}
#[test]
fn tenantstatus_activating_strum() {
// tests added, because we use these for metrics
let examples = [
(line!(), TenantState::Loading, "Loading"),
(line!(), TenantState::Attaching, "Attaching"),
(
line!(),
TenantState::Activating(ActivatingFrom::Loading),
"Activating",
),
(
line!(),
TenantState::Activating(ActivatingFrom::Attaching),
"Activating",
),
(line!(), TenantState::Active, "Active"),
(line!(), TenantState::Stopping, "Stopping"),
(
line!(),
TenantState::Broken {
reason: "Example".into(),
backtrace: "Looooong backtrace".into(),
},
"Broken",
),
];
for (line, rendered, expected) in examples {
let actual: &'static str = rendered.into();
assert_eq!(actual, expected, "example on {line}");
}
}
}

View File

@@ -25,8 +25,6 @@ utils.workspace = true
pin-project-lite.workspace = true
workspace_hack.workspace = true
leaky-bucket = "1.0"
[dev-dependencies]
tempfile.workspace = true
test-context.workspace = true

View File

@@ -37,8 +37,6 @@ pub const DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS: u32 = 10;
/// https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/UsingWithRDS.IAMDBAuth.html
/// ~3500 PUT/COPY/POST/DELETE or 5500 GET/HEAD S3 requests
/// https://aws.amazon.com/premiumsupport/knowledge-center/s3-request-limit-avoid-throttling/
///
/// IAM ratelimit should never be observed with caching credentials provider.
pub const DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT: usize = 100;
/// No limits on the client side, which currenltly means 1000 for AWS S3.
/// https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax

View File

@@ -21,7 +21,10 @@ use aws_sdk_s3::{
};
use aws_smithy_http::body::SdkBody;
use hyper::Body;
use tokio::io;
use tokio::{
io::{self, AsyncRead},
sync::Semaphore,
};
use tokio_util::io::ReaderStream;
use tracing::debug;
@@ -102,8 +105,9 @@ pub struct S3Bucket {
prefix_in_bucket: Option<String>,
max_keys_per_list_response: Option<i32>,
// Every request to S3 can be throttled or cancelled, if a certain number of requests per second is exceeded.
// Same goes to IAM, which is queried before every S3 request, if enabled. IAM has even lower RPS threshold.
// The helps to ensure we don't exceed the thresholds.
concurrency_limiter: Arc<leaky_bucket::RateLimiter>,
concurrency_limiter: Arc<Semaphore>,
}
#[derive(Default)]
@@ -154,24 +158,12 @@ impl S3Bucket {
}
prefix
});
let rps = aws_config.concurrency_limit.get();
let concurrency_limiter = leaky_bucket::RateLimiter::builder()
.max(rps)
.initial(0)
// refill it by rps every second. this means the (rps+1)th request will have to wait for
// 1 second from earliest.
.refill(rps)
.interval(std::time::Duration::from_secs(1))
.fair(true)
.build();
Ok(Self {
client,
bucket_name: aws_config.bucket_name.clone(),
max_keys_per_list_response: aws_config.max_keys_per_list_response,
prefix_in_bucket,
concurrency_limiter: Arc::new(concurrency_limiter),
concurrency_limiter: Arc::new(Semaphore::new(aws_config.concurrency_limit.get())),
})
}
@@ -203,10 +195,13 @@ impl S3Bucket {
}
async fn download_object(&self, request: GetObjectRequest) -> Result<Download, DownloadError> {
// while the download could take a long time with `leaky_bucket` we have nothing to release
// once the download is done. this is because with "requests per second" rate limiting on
// s3, there should be no meaning for the long requests.
self.concurrency_limiter.clone().acquire_owned(1).await;
let permit = self
.concurrency_limiter
.clone()
.acquire_owned()
.await
.context("Concurrency limiter semaphore got closed during S3 download")
.map_err(DownloadError::Other)?;
metrics::inc_get_object();
@@ -224,9 +219,10 @@ impl S3Bucket {
let metadata = object_output.metadata().cloned().map(StorageMetadata);
Ok(Download {
metadata,
download_stream: Box::pin(io::BufReader::new(
download_stream: Box::pin(io::BufReader::new(RatelimitedAsyncRead::new(
permit,
object_output.body.into_async_read(),
)),
))),
})
}
Err(SdkError::ServiceError(e)) if matches!(e.err(), GetObjectError::NoSuchKey(_)) => {
@@ -242,6 +238,32 @@ impl S3Bucket {
}
}
pin_project_lite::pin_project! {
/// An `AsyncRead` adapter which carries a permit for the lifetime of the value.
struct RatelimitedAsyncRead<S> {
permit: tokio::sync::OwnedSemaphorePermit,
#[pin]
inner: S,
}
}
impl<S: AsyncRead> RatelimitedAsyncRead<S> {
fn new(permit: tokio::sync::OwnedSemaphorePermit, inner: S) -> Self {
RatelimitedAsyncRead { permit, inner }
}
}
impl<S: AsyncRead> AsyncRead for RatelimitedAsyncRead<S> {
fn poll_read(
self: std::pin::Pin<&mut Self>,
cx: &mut std::task::Context<'_>,
buf: &mut io::ReadBuf<'_>,
) -> std::task::Poll<std::io::Result<()>> {
let this = self.project();
this.inner.poll_read(cx, buf)
}
}
#[async_trait::async_trait]
impl RemoteStorage for S3Bucket {
/// See the doc for `RemoteStorage::list_prefixes`
@@ -267,7 +289,12 @@ impl RemoteStorage for S3Bucket {
let mut continuation_token = None;
loop {
self.concurrency_limiter.acquire_one().await;
let _guard = self
.concurrency_limiter
.acquire()
.await
.context("Concurrency limiter semaphore got closed during S3 list")
.map_err(DownloadError::Other)?;
metrics::inc_list_objects();
@@ -312,9 +339,11 @@ impl RemoteStorage for S3Bucket {
to: &RemotePath,
metadata: Option<StorageMetadata>,
) -> anyhow::Result<()> {
// similarly to downloads, the permit does not have live through the upload, but instead we
// are rate limiting requests per second.
self.concurrency_limiter.acquire_one().await;
let _guard = self
.concurrency_limiter
.acquire()
.await
.context("Concurrency limiter semaphore got closed during S3 upload")?;
metrics::inc_put_object();
@@ -369,7 +398,11 @@ impl RemoteStorage for S3Bucket {
}
async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
self.concurrency_limiter.acquire_one().await;
let _guard = self
.concurrency_limiter
.acquire()
.await
.context("Concurrency limiter semaphore got closed during S3 delete")?;
metrics::inc_delete_object();

View File

@@ -0,0 +1,33 @@
use std::sync::Arc;
use tokio::sync::{mpsc, Mutex};
/// While a reference is kept around, the associated [`Barrier::wait`] will wait.
///
/// Can be cloned, moved and kept around in futures as "guard objects".
#[derive(Clone)]
pub struct Completion(mpsc::Sender<()>);
/// Barrier will wait until all clones of [`Completion`] have been dropped.
#[derive(Clone)]
pub struct Barrier(Arc<Mutex<mpsc::Receiver<()>>>);
impl Barrier {
pub async fn wait(self) {
self.0.lock().await.recv().await;
}
pub async fn maybe_wait(barrier: Option<Barrier>) {
if let Some(b) = barrier {
b.wait().await
}
}
}
/// Create new Guard and Barrier pair.
pub fn channel() -> (Completion, Barrier) {
let (tx, rx) = mpsc::channel::<()>(1);
let rx = Mutex::new(rx);
let rx = Arc::new(rx);
(Completion(tx), Barrier(rx))
}

View File

@@ -1,5 +1,5 @@
use crate::auth::{Claims, JwtAuth};
use crate::http::error;
use crate::http::error::{api_error_handler, route_error_handler, ApiError};
use anyhow::{anyhow, Context};
use hyper::header::{HeaderName, AUTHORIZATION};
use hyper::http::HeaderValue;
@@ -16,8 +16,6 @@ use std::future::Future;
use std::net::TcpListener;
use std::str::FromStr;
use super::error::ApiError;
static SERVE_METRICS_COUNT: Lazy<IntCounter> = Lazy::new(|| {
register_int_counter!(
"libmetrics_metric_handler_requests_total",
@@ -35,8 +33,12 @@ struct RequestId(String);
/// Adds a tracing info_span! instrumentation around the handler events,
/// logs the request start and end events for non-GET requests and non-200 responses.
///
/// Usage: Replace `my_handler` with `|r| request_span(r, my_handler)`
///
/// Use this to distinguish between logs of different HTTP requests: every request handler wrapped
/// in this type will get request info logged in the wrapping span, including the unique request ID.
/// with this will get request info logged in the wrapping span, including the unique request ID.
///
/// This also handles errors, logging them and converting them to an HTTP error response.
///
/// There could be other ways to implement similar functionality:
///
@@ -54,60 +56,56 @@ struct RequestId(String);
/// tries to achive with its `.instrument` used in the current approach.
///
/// If needed, a declarative macro to substitute the |r| ... closure boilerplate could be introduced.
pub struct RequestSpan<E, R, H>(pub H)
pub async fn request_span<R, H>(request: Request<Body>, handler: H) -> R::Output
where
E: Into<Box<dyn std::error::Error + Send + Sync>> + 'static,
R: Future<Output = Result<Response<Body>, E>> + Send + 'static,
H: Fn(Request<Body>) -> R + Send + Sync + 'static;
impl<E, R, H> RequestSpan<E, R, H>
where
E: Into<Box<dyn std::error::Error + Send + Sync>> + 'static,
R: Future<Output = Result<Response<Body>, E>> + Send + 'static,
H: Fn(Request<Body>) -> R + Send + Sync + 'static,
R: Future<Output = Result<Response<Body>, ApiError>> + Send + 'static,
H: FnOnce(Request<Body>) -> R + Send + Sync + 'static,
{
/// Creates a tracing span around inner request handler and executes the request handler in the contex of that span.
/// Use as `|r| RequestSpan(my_handler).handle(r)` instead of `my_handler` as the request handler to get the span enabled.
pub async fn handle(self, request: Request<Body>) -> Result<Response<Body>, E> {
let request_id = request.context::<RequestId>().unwrap_or_default().0;
let method = request.method();
let path = request.uri().path();
let request_span = info_span!("request", %method, %path, %request_id);
let request_id = request.context::<RequestId>().unwrap_or_default().0;
let method = request.method();
let path = request.uri().path();
let request_span = info_span!("request", %method, %path, %request_id);
let log_quietly = method == Method::GET;
async move {
let cancellation_guard = RequestCancelled::warn_when_dropped_without_responding();
if log_quietly {
debug!("Handling request");
} else {
info!("Handling request");
}
// Note that we reuse `error::handler` here and not returning and error at all,
// yet cannot use `!` directly in the method signature due to `routerify::RouterBuilder` limitation.
// Usage of the error handler also means that we expect only the `ApiError` errors to be raised in this call.
//
// Panics are not handled separately, there's a `tracing_panic_hook` from another module to do that globally.
let res = (self.0)(request).await;
cancellation_guard.disarm();
match res {
Ok(response) => {
let response_status = response.status();
if log_quietly && response_status.is_success() {
debug!("Request handled, status: {response_status}");
} else {
info!("Request handled, status: {response_status}");
}
Ok(response)
}
Err(e) => Ok(error::handler(e.into()).await),
}
let log_quietly = method == Method::GET;
async move {
let cancellation_guard = RequestCancelled::warn_when_dropped_without_responding();
if log_quietly {
debug!("Handling request");
} else {
info!("Handling request");
}
// No special handling for panics here. There's a `tracing_panic_hook` from another
// module to do that globally.
let res = handler(request).await;
cancellation_guard.disarm();
// Log the result if needed.
//
// We also convert any errors into an Ok response with HTTP error code here.
// `make_router` sets a last-resort error handler that would do the same, but
// we prefer to do it here, before we exit the request span, so that the error
// is still logged with the span.
//
// (Because we convert errors to Ok response, we never actually return an error,
// and we could declare the function to return the never type (`!`). However,
// using `routerify::RouterBuilder` requires a proper error type.)
match res {
Ok(response) => {
let response_status = response.status();
if log_quietly && response_status.is_success() {
debug!("Request handled, status: {response_status}");
} else {
info!("Request handled, status: {response_status}");
}
Ok(response)
}
Err(err) => Ok(api_error_handler(err)),
}
.instrument(request_span)
.await
}
.instrument(request_span)
.await
}
/// Drop guard to WARN in case the request was dropped before completion.
@@ -207,10 +205,8 @@ pub fn make_router() -> RouterBuilder<hyper::Body, ApiError> {
.middleware(Middleware::post_with_info(
add_request_id_header_to_response,
))
.get("/metrics", |r| {
RequestSpan(prometheus_metrics_handler).handle(r)
})
.err_handler(error::handler)
.get("/metrics", |r| request_span(r, prometheus_metrics_handler))
.err_handler(route_error_handler)
}
pub fn attach_openapi_ui(
@@ -220,12 +216,14 @@ pub fn attach_openapi_ui(
ui_mount_path: &'static str,
) -> RouterBuilder<hyper::Body, ApiError> {
router_builder
.get(spec_mount_path, move |r| {
RequestSpan(move |_| async move { Ok(Response::builder().body(Body::from(spec)).unwrap()) })
.handle(r)
})
.get(ui_mount_path, move |r| RequestSpan( move |_| async move {
Ok(Response::builder().body(Body::from(format!(r#"
.get(spec_mount_path,
move |r| request_span(r, move |_| async move {
Ok(Response::builder().body(Body::from(spec)).unwrap())
})
)
.get(ui_mount_path,
move |r| request_span(r, move |_| async move {
Ok(Response::builder().body(Body::from(format!(r#"
<!DOCTYPE html>
<html lang="en">
<head>
@@ -255,7 +253,8 @@ pub fn attach_openapi_ui(
</body>
</html>
"#, spec_mount_path))).unwrap())
}).handle(r))
})
)
}
fn parse_token(header_value: &str) -> Result<&str, ApiError> {

View File

@@ -83,13 +83,24 @@ impl HttpErrorBody {
}
}
pub async fn handler(err: routerify::RouteError) -> Response<Body> {
let api_error = err
.downcast::<ApiError>()
.expect("handler should always return api error");
pub async fn route_error_handler(err: routerify::RouteError) -> Response<Body> {
match err.downcast::<ApiError>() {
Ok(api_error) => api_error_handler(*api_error),
Err(other_error) => {
// We expect all the request handlers to return an ApiError, so this should
// not be reached. But just in case.
error!("Error processing HTTP request: {other_error:?}");
HttpErrorBody::response_from_msg_and_status(
other_error.to_string(),
StatusCode::INTERNAL_SERVER_ERROR,
)
}
}
}
pub fn api_error_handler(api_error: ApiError) -> Response<Body> {
// Print a stack trace for Internal Server errors
if let ApiError::InternalServerError(_) = api_error.as_ref() {
if let ApiError::InternalServerError(_) = api_error {
error!("Error processing HTTP request: {api_error:?}");
} else {
error!("Error processing HTTP request: {api_error:#}");

View File

@@ -60,6 +60,9 @@ pub mod tracing_span_assert;
pub mod rate_limit;
/// Simple once-barrier and a guard which keeps barrier awaiting.
pub mod completion;
mod failpoint_macro_helpers {
/// use with fail::cfg("$name", "return(2000)")