New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Async IO with CLibvenice #15

Open

robertjpayne wants to merge 6 commits into master from async

robertjpayne commented Dec 1, 2016

This PR makes PostgreSQL non-blocking and play nice with Libvenice:

Uses async connection
Uses async querying

@Danappelxx please review 👍

robertjpayne added 4 commits

December 1, 2016 22:47


          Make PostgreSQL use libmill for non-blocking goodness

c47bd9b


          Add safety around open/close

0eb0a9b


          Tidy up Postgres Connect to be async

d053c79


          Remove unecessary print

eda562f

Danappelxx requested changes

View reviewed changes

Contributor

Danappelxx left a comment

This is really great! Sorry took so long to review. The PR needs a bit of cleanup (not a problem), but I'm not 100% certain the semantics are correct. More in the next comment.

Sources/PostgreSQL/Connection.swift

+                          case PGRES_POLLING_OK:
+                              break loop
+                          case PGRES_POLLING_READING:
+                              mill_fdwait_(fd, FDW_IN, 15.seconds.fromNow().int64milliseconds, nil)

Contributor

Danappelxx Dec 1, 2016

You can use Venice.poll here

Author

robertjpayne Dec 17, 2016

Venice.poll doesn't clean the file descriptor when it's done

Sources/PostgreSQL/Connection.swift

+                              mill_fdwait_(fd, FDW_IN, 15.seconds.fromNow().int64milliseconds, nil)
+                              mill_fdclean_(fd)
+                          case PGRES_POLLING_WRITING:
+                              mill_fdwait_(fd, FDW_OUT, 15.seconds.fromNow().int64milliseconds, nil)

Contributor

Danappelxx Dec 1, 2016

Here too

Author

robertjpayne Dec 17, 2016

Venice.poll doesn't clean the file descriptor when it's done

Sources/PostgreSQL/Connection.swift

		import CLibpq
		import CLibvenice

Contributor

Danappelxx Dec 1, 2016

CLibvenice -> Venice - I don't think we're using any api's that venice doesn't wrap (if we are, should use both?)

Author

robertjpayne Dec 17, 2016

Does Venice have an API to explicitly clean up a fdwait? We have to wait and then immediately clean it up otherwise when libpq goes to read it causes a crash.

Contributor

Danappelxx Dec 17, 2016

Hmm, doesn't look like it. Venice.poll just wraps fdwait - I guess we can use fdclean from CLibvenice.

Sources/PostgreSQL/Connection.swift

-                      return try Result(result)
+              extension Collection where Iterator.Element == String {
+                  func withUnsafeCStringArray<T>(_ body: (UnsafePointer<UnsafePointer<Int8>?>) throws -> T) rethrows -> T {

Contributor

Danappelxx Dec 1, 2016

I think you forgot to use this :)

Author

robertjpayne Dec 17, 2016

Haha yes! I need to get rid of it 👍

Sources/PostgreSQL/Connection.swift

+                          throw ConnectionError(description: "Connection already opened.")
+                      }
+                      var components = URLComponents()

Contributor

Danappelxx Dec 17, 2016

IMO, better style would be

let url: String = {
    var components = URLComponents()
    // ...
    return components.url!.absoluteString
}()

Author

robertjpayne Dec 17, 2016

Personally just think that is a nit, it's a local scope var and mutated either way step by step…

Sources/PostgreSQL/Connection.swift

+                          throw mostRecentError ?? ConnectionError(description: "Could not get file descriptor.")
+                      }
+                      loop: while true {

Contributor

Danappelxx Dec 17, 2016

Loop isn't nested, so I don't think a label is necessary

Author

robertjpayne Dec 17, 2016

Yeah I can get rid of that loop label

Author

robertjpayne Dec 17, 2016

Actually you can't break from a switch to break the loop, hence the label?

Sources/PostgreSQL/Connection.swift

+                              mill_fdwait_(fd, FDW_OUT, 15.seconds.fromNow().int64milliseconds, nil)
+                              mill_fdclean_(fd)
+                          case PGRES_POLLING_ACTIVE:
+                              break

Contributor

Danappelxx Dec 17, 2016

continue instead of break to avoid the label. This might be "clever", though...

Author

robertjpayne Dec 17, 2016

Only if the loop label can be fixed

Sources/PostgreSQL/Connection.swift

                       var parameterData = [UnsafePointer<Int8>?]()
                       var deallocators = [() -> ()]()
                       defer { deallocators.forEach { $0() } }
-                      for parameter in parameters {
+                      if let parameters = parameters {

Contributor

Danappelxx Dec 17, 2016

this function got really long - needs some cleaning

Author

robertjpayne Dec 17, 2016

Possibly, it works though? It's still pretty small

Tests/PostgreSQLTests/PostgreSQLTests.swift

		do {
		connection = try! PostgreSQL.Connection(info: .init(URL(string: "postgres://localhost:5432/swift_test")!))

Contributor

Danappelxx Dec 17, 2016

Is there a reason the connection is being reconstructed each time? Is this the right way to use the api now?

Author

robertjpayne Dec 17, 2016

For tests each test should use it's own connection, we shouldn't share it due to busy states/dangling etc…

Connections are also unrecoverable ( you can't "reconnect" )

Contributor

Danappelxx commented Dec 17, 2016

Two things lead me to think that we're not polling quite right:

In my testing, the blocking version was actually faster than this nonblocking version. In theory, this is not supposed to happen. I tested using wrk -d 10 -t 10 -c 100 http://localhost:8080/api/v1/posts on an endpoint which is 100% database io.

without this pr:

> wrk -d 10 -t 10 -c 100 http://localhost:8080/api/v1/posts
Running 10s test @ http://localhost:8080/api/v1/posts
  10 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    44.86ms    3.71ms 101.55ms   87.31%
    Req/Sec   223.54     22.24   300.00     75.20%
  22326 requests in 10.05s, 8.84MB read
Requests/sec:   2221.08
Transfer/sec:      0.88MB

with this pr:

> wrk -d 10 -t 10 -c 100 http://localhost:8080/api/v1/posts
Running 10s test @ http://localhost:8080/api/v1/posts
  10 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     3.06ms    5.66ms  82.92ms   87.77%
    Req/Sec   670.91    280.70     1.16k    66.05%
  10877 requests in 10.08s, 3.56MB read
  Socket errors: connect 0, read 2360, write 0, timeout 0
  Non-2xx or 3xx responses: 2459
Requests/sec:   1079.14
Transfer/sec:    361.59KB

In the logs for this test, I found 2425 occurrences of Error: another command is already in progress. This is probably a symptom of the same issue :)

Contributor

Danappelxx commented Dec 17, 2016 •

edited

Loading

Also, it didn't actually build for me (3.0.2). Are you using a modified version of libmill/venice? mill_fdclean_ doesn't exist in libvenice (it's just fdclean). Renaming a few function calls fixed it.

Author

robertjpayne commented Dec 17, 2016

@Danappelxx as per the performance your tests are re-using a connection I assume? This is why it appears to perform slower.

Before a single request will execute the full SQL command before any other request in the entire process could continue thus you sort of have an implicit lock on any SQL execution.

Now when libpq is waiting for IO another coroutine will pick up a new request and start to try and use the same Postgres connection, Postgres is half way through executing another SQL statement and hence you get the errors.

For this async stuff to work we really need a connection pool that a request checks a connection out and when finished it puts it back.


          Make it build with swift 3.0.2

912fe0d

Contributor

Danappelxx commented Dec 23, 2016

Did some more testing and I'm satisfied! Works great with a connection pool as you suggested. Just a little bit of cleanup and its good to go.


          Remove unused code

7f6b3cf

Danappelxx approved these changes

View reviewed changes

Contributor

Danappelxx commented Dec 23, 2016

(merge whenever you're ready)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet