Giters

davepacheco / async-bb8-cancel

demos a failure mode involving async cancellation

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Task cancellation failure mode

This repo contains a relatively simple program demonstrating a particular failure mode resulting from cancel-unsafe code.

The program here is basically just an HTTP server sitting in front of a PostgreSQL/CockroachDB database. The HTTP server exposes endpoints for fetching and bumping a counter that’s stored in the database:

  • POST /reset: deletes the existing database table (if any) and creates the table anew with an initial counter value of 1

  • GET /counter: fetches the current value of the counter in the database

  • POST /bump: bumps the current value of the counter in the database

  • GET /sleep: opens a database transaction and then goes to sleep for 5 seconds

Prerequisites

The program assumes it can start an HTTP server on 127.0.0.1 port 12344.

The program assumes that a PostgreSQL-compatible database is already running with a database called "defaultdb" and that it can be accessed without authentication using the URL postgresql://root@127.0.0.1:12345/defaultdb?sslmode=disable. If you’re working on Omicron, you can start a temporary database instance using cargo run --bin=omicron-dev db-run --listen-port=12345 --no-populate (or just omicron-dev db-run --listen-port-12345 --no-populate).

The program logs in bunyan format to stdout. You can use the standard bunyan tool to view the log. In the examples below I use looker instead.

The program links to libpq, so you have to have that installed. See "Installing Prerequisites" in the Omicron documentation. The build process for this program does not provide the runtime linker flags required to set RPATH in the generated binary to point to libpq. On your system, you may need to specify LD_LIBRARY_PATH.

To make the output more readable, I set the prompt with export PS1='\n$ '.

Basic usage

Let’s show how this works first. I’ve got the database started as shown above. Now I’ll start this program:

$ LD_LIBRARY_PATH=/opt/ooce/pgsql-13/lib/amd64 cargo run | looker
note: configured to log to "/dev/stdout"
00:03:50.124Z INFO cancel-repro: setting up dropshot server
00:03:50.124Z DEBG cancel-repro: registered endpoint
    local_addr = 127.0.0.1:12344
    method = POST
    path = /bump
00:03:50.124Z DEBG cancel-repro: registered endpoint
    local_addr = 127.0.0.1:12344
    method = GET
    path = /counter
00:03:50.124Z DEBG cancel-repro: registered endpoint
    local_addr = 127.0.0.1:12344
    method = POST
    path = /reset
00:03:50.124Z DEBG cancel-repro: registered endpoint
    local_addr = 127.0.0.1:12344
    method = GET
    path = /sleep
00:03:50.124Z INFO cancel-repro: listening
    local_addr = 127.0.0.1:12344
00:03:50.124Z DEBG cancel-repro: DTrace USDT probes compiled out, not registering
    local_addr = 127.0.0.1:12344
00:03:50.124Z INFO cancel-repro: set up dropshot server
    local_address = 127.0.0.1:12344

First, let’s fetch the counter value:

$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 500 Internal Server Error
content-type: application/json
x-request-id: 07bb02a2-cbdb-48ce-b910-89c0adc09cbd
content-length: 283
date: Thu, 22 Jun 2023 00:06:04 GMT

{
  "request_id": "07bb02a2-cbdb-48ce-b910-89c0adc09cbd",
  "message": "loading counter: Failure accessing a connection: Failed to issue a query: relation \"counter\" does not exist: Failed to issue a query: relation \"counter\" does not exist: relation \"counter\" does not exist"
}

Oops! We forgot to initialize the database. Let’s do that:

curl -i http://127.0.0.1:12344/reset -X POST
HTTP/1.1 200 OK
content-type: application/json
x-request-id: 736e6470-5753-4236-9df6-5f6a6a037459
content-length: 4
date: Thu, 22 Jun 2023 00:06:37 GMT

"ok"

Now fetch the counter:

$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: 115cfe48-3e3c-4043-bf24-c0600c45ab02
content-length: 1
date: Thu, 22 Jun 2023 00:06:49 GMT

1

Let’s be sure:

$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: 77bc4855-98f5-46a1-9c9e-e8a05921ad93
content-length: 1
date: Thu, 22 Jun 2023 00:07:01 GMT

1

Great. Let’s bump it:

$ curl -i http://127.0.0.1:12344/bump -X POST
HTTP/1.1 204 No Content
x-request-id: 6e89799a-23df-4acf-b699-f3cb6a0ddc15
date: Thu, 22 Jun 2023 00:07:16 GMT
$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: 80e83d06-a053-4a62-a0c5-40fa3758eac1
content-length: 1
date: Thu, 22 Jun 2023 00:07:18 GMT

2

Great. Let’s bump it a few more times and fetch it a few more:

$ curl -i http://127.0.0.1:12344/bump -X POST
HTTP/1.1 204 No Content
x-request-id: 2680cdd9-1011-4305-a01b-41ae0c2a0f8b
date: Thu, 22 Jun 2023 00:08:01 GMT

$ curl -i http://127.0.0.1:12344/bump -X POST
HTTP/1.1 204 No Content
x-request-id: 17d2afbb-29f3-4ccf-ac6d-94ce3fe2ae2e
date: Thu, 22 Jun 2023 00:08:01 GMT

$ curl -i http://127.0.0.1:12344/bump -X POST
HTTP/1.1 204 No Content
x-request-id: b9d35edd-ddd6-4fbe-b205-c4498d222b86
date: Thu, 22 Jun 2023 00:08:02 GMT

$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: c57e6b7d-630c-4a62-80b9-c2f4e55484cf
content-length: 1
date: Thu, 22 Jun 2023 00:08:04 GMT

5
$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: 3e557ff6-3cf0-404d-8cb0-e28b04f5cf19
content-length: 1
date: Thu, 22 Jun 2023 00:08:04 GMT

5

Now, we can stop our server and start it again:

^C
dap@ivanova cancel-repro $ LD_LIBRARY_PATH=/opt/ooce/pgsql-13/lib/amd64 cargo run | looker
    Finished dev [unoptimized + debuginfo] target(s) in 0.10s
     Running `target/debug/cancel-repro`
note: configured to log to "/dev/stdout"
00:09:20.271Z INFO cancel-repro: setting up dropshot server
00:09:20.271Z DEBG cancel-repro: registered endpoint
    local_addr = 127.0.0.1:12344
    method = POST
    path = /bump
00:09:20.272Z DEBG cancel-repro: registered endpoint
    local_addr = 127.0.0.1:12344
    method = GET
    path = /counter
00:09:20.272Z DEBG cancel-repro: registered endpoint
    local_addr = 127.0.0.1:12344
    method = POST
    path = /reset
00:09:20.272Z DEBG cancel-repro: registered endpoint
    local_addr = 127.0.0.1:12344
    method = GET
    path = /sleep
00:09:20.272Z INFO cancel-repro: listening
    local_addr = 127.0.0.1:12344
00:09:20.272Z DEBG cancel-repro: DTrace USDT probes compiled out, not registering
    local_addr = 127.0.0.1:12344
00:09:20.272Z INFO cancel-repro: set up dropshot server
    local_address = 127.0.0.1:12344

and of course the counter value is still the same, since it was stored in the database:

$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: 1cff3fbd-b4c9-4297-a8c6-615af7b3d75a
content-length: 1
date: Thu, 22 Jun 2023 00:10:02 GMT

5

and of course we can bump it:

$ curl -i http://127.0.0.1:12344/bump -X POST
HTTP/1.1 204 No Content
x-request-id: 66f00e76-3d28-4317-b41e-2e5b19b7a7cc
date: Thu, 22 Jun 2023 00:10:08 GMT


$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: fe7adac4-aea7-4c6f-862f-ed4d62b82e36
content-length: 1
date: Thu, 22 Jun 2023 00:10:09 GMT

6

There’s also a "sleep" endpoint that will just sleep for 5 seconds. We can hit that and do all the same stuff we did before:

$ curl -i http://127.0.0.1:12344/sleep
HTTP/1.1 200 OK
content-type: application/json
x-request-id: 9a7717f3-78ca-41bc-a4c2-79d5cb20273b
content-length: 4
date: Thu, 22 Jun 2023 00:11:28 GMT

"ok"
$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: 258d5e3c-eec3-4614-91b8-56219ca63776
content-length: 1
date: Thu, 22 Jun 2023 00:11:30 GMT

6
$ curl -i http://127.0.0.1:12344/bump -X POST
HTTP/1.1 204 No Content
x-request-id: 32b99dec-6281-471d-8396-b8239fa7ff39
date: Thu, 22 Jun 2023 00:11:33 GMT


$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: 02b1826d-5a83-475c-81c5-f84ee7aa9e12
content-length: 1
date: Thu, 22 Jun 2023 00:11:35 GMT

7

Making things interesting

But what happens if we cancel the sleep endpoint? All this endpoint does is open a database transaction and then call tokio::time::sleep for 5 seconds. What do you think happens?

$ curl -i http://127.0.0.1:12344/sleep
^C

There’s also logic in the endpoint to notice when it’s been cancelled and report it. So we get this log message:

00:12:41.786Z ERRO cancel-repro: api_sleep() cancelled
    local_addr = 127.0.0.1:12344
    method = GET
    remote_addr = 127.0.0.1:62550
    req_id = bba4c1f3-59cb-452e-b324-9a99d65569b4
    uri = /sleep

Things appear to be working just fine:

$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: 58630d86-e3b7-4a58-93b6-8c88492be733
content-length: 1
date: Thu, 22 Jun 2023 00:13:37 GMT

7
$ curl -i http://127.0.0.1:12344/bump -X POST
HTTP/1.1 204 No Content
x-request-id: c39523c0-2cf0-4de5-8cd4-ff40f84a7494
date: Thu, 22 Jun 2023 00:13:39 GMT


$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: a0bf588f-c64e-407a-9442-de7e6bd73fad
content-length: 1
date: Thu, 22 Jun 2023 00:13:40 GMT

8
$ curl -i http://127.0.0.1:12344/bump -X POST
HTTP/1.1 204 No Content
x-request-id: 0ab3363e-5012-49f1-83cd-e3ae9c7d4d4d
date: Thu, 22 Jun 2023 00:13:42 GMT


$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: 539bab87-1564-4378-8127-14b1f8d85d86
content-length: 1
date: Thu, 22 Jun 2023 00:13:43 GMT

9

Let’s shut down the server and start it up again:

...
00:13:43.510Z INFO cancel-repro: request completed
    local_addr = 127.0.0.1:12344
    method = GET
    remote_addr = 127.0.0.1:61645
    req_id = 539bab87-1564-4378-8127-14b1f8d85d86
    response_code = 200
    uri = /counter
^C
$ LD_LIBRARY_PATH=/opt/ooce/pgsql-13/lib/amd64 cargo run | looker
    Finished dev [unoptimized + debuginfo] target(s) in 0.09s
     Running `target/debug/cancel-repro`
note: configured to log to "/dev/stdout"
00:14:21.851Z INFO cancel-repro: setting up dropshot server
00:14:21.852Z DEBG cancel-repro: registered endpoint
    local_addr = 127.0.0.1:12344
    method = POST
    path = /bump
00:14:21.852Z DEBG cancel-repro: registered endpoint
    local_addr = 127.0.0.1:12344
    method = GET
    path = /counter
00:14:21.852Z DEBG cancel-repro: registered endpoint
    local_addr = 127.0.0.1:12344
    method = POST
    path = /reset
00:14:21.852Z DEBG cancel-repro: registered endpoint
    local_addr = 127.0.0.1:12344
    method = GET
    path = /sleep
00:14:21.852Z INFO cancel-repro: listening
    local_addr = 127.0.0.1:12344
00:14:21.852Z DEBG cancel-repro: DTrace USDT probes compiled out, not registering
    local_addr = 127.0.0.1:12344
00:14:21.852Z INFO cancel-repro: set up dropshot server
    local_address = 127.0.0.1:12344
...

and the counter will be:

$ curl -i http://127.0.0.1:12344/counter
HTTP/1.1 200 OK
content-type: application/json
x-request-id: 8dbca33c-fcfc-47fc-a869-65488657a1f9
content-length: 1
date: Thu, 22 Jun 2023 00:15:03 GMT

7

What?! The counter was 9 before the restart!

What happened?

async-bb8-diesel provides a connection.transaction_async(closure) function that, just like its synchronous counterpart in bb8-diesel, opens a transaction, invokes the closure, and then commits or rolls back the transaction based on the result. The problem is that transaction_async is not cancel-safe. It has to await on your closure. But if it gets cancelled at that point, then it never commits or rolls back the transaction. The connection gets put back into the connection pool.

The consequence of this is that when any subsequent operation grabs this database connection, it will unexpectedly be running inside a database transaction. That transaction will never be committed because we’ve lost track of the fact that we’re in a transaction. What’s really rough about this failure mode is that a subsequent operation can continue successfully executing SQL statements, making updates and fetching data, and everything will appear to be working. But it’s all provisional on that transaction committing.

Back to the example: when we cancel the "sleep" API call, we triggered this bug. Every subsequent operation we did to fetch and bump the counter was inside that database transaction and reporting a view of the world in the alternate reality where that transaction commits. When we shut down and started our server again, we shut down the connection. (The database would have implicitly rolled back that transaction, but that doesn’t even matter.) Once the server came up again and we fetched the counter value, we were grabbing the live value from the database, not the one visible to that ill-fated transaction. So the counter appeared to go backwards.

To make this problem more obvious, this program configures a database pool with only one connection so that we always hit the problem. In a more realistic system, the server would have multiple database connections. If you fetched the counter value after hitting the bug, you’d get the right value much of the time! Similarly, most of the "bump" operations would work and actually update the database. It’s just whatever operations got the unlucky connection where (1) updates are completing successfully but being ignored and (2) they will continue read back whatever state has been updated in that connection.

About

demos a failure mode involving async cancellation

Languages

Language:Rust 98.2%Language:Shell 1.8%