Skip to content

Latest commit

 

History

History
174 lines (116 loc) · 7.5 KB

internals.md

File metadata and controls

174 lines (116 loc) · 7.5 KB

Odyssey architecture and internals

Odyssey heavily depends on two libraries, which were originally created during its development: Machinarium and Kiwi.

Machinarium

Machinarium extensively used for organization of multi-thread processing, cooperative multi-tasking and networking IO. All Odyssey threads are run in context of machinarium machines - pthreads with coroutine schedulers placed on top of epoll(7) event loop.

Odyssey does not directly use or create multi-tasking primitives such as OS threads and mutexes. All synchronization is done using message passing and transparently handled by machinarium.

Repository: third_party/machinarium

Kiwi

Kiwi provides functions for constructing, reading and validating PostgreSQL protocol requests messages. By design, all PostgreSQL specific details should be provided by Kiwi library.

Repository: third_party/kiwi

Core components

                                              main()
                                           .----------.
                                           | instance |
                           thread          '----------'
                         .--------.                          .-------------.
                         | system |                          | worker_pool |
                         '--------'                          '-------------'
                  .--------.    .---------.           .---------.         .---------.
                  | router |    | servers |           | worker0 |   ...   | workerN |
                  '--------'    '---------'           '---------'         '---------'
                  .---------.    .------.               thread              thread
                  | console |    | cron |
                  '---------'    '------'

Instance

Application entry point.

Handle initialization. Read configuration file, prepare loggers, parse cli options. Run system and worker_pool threads.

sources/instance.h, sources/instance.c

System

Start router, cron and console subsystems.

Create listen server one for each resolved address. Each listen server runs inside own coroutine. Server coroutine mostly waits on machine_accept().

On incoming connection, new client context is created and notification message is sent to next worker using workerpool_feed(). Client IO context is not attached to any epoll(7) context yet.

Handle signals using machine_signal_wait(). On SIGHUP: do versional config reload, add new databases and obsolete old ones. On SIGUSR1: reopen log file. On SIGUSR2: graceful shutdown. On SIGINT, SIGTERM: call exit(3). Other threads are blocked from receiving signals.

sources/system.h, sources/system.c

Router

Handle client registration and routing requests. Do client-to-server attachment and detachment. Ensure connection limits and client pool queueing. Handle implicit Cancel client request, since access to server pool is required to match a client key.

Router works in request-reply manner: client (from worker thread) sends a request message to router and waits for reply. Could be a potential hot spot (not an issue at the moment).

sources/router.h, sources/router.c

Cron

Do periodic service tasks, like idle server connection expiration and database config obsoletion.

sources/cron.h, sources/cron.c

Worker and worker pool

Worker thread (machinarium machine) waits on incoming connection notification queue. On new connection event, create new frontend coroutine and handle client (frontend) lifecycle. Each worker thread can host thousands of client coroutines.

Worker pool is responsible for maintaining a thread pool of workers. Threads are machinarium machines, created using machine_create().

sources/worker.h, sources/worker.c, sources/worker_pool.h, sources/worker_pool.c

Single worker mode

To reduce multi-thread communication overhead, Odyssey handles case with a single worker (workers 1) differently.

Instead of creating separate thread + coroutine for each worker, only one worker coroutine created inside system thread. All message channels machine_channel_create() created marked as non-shared. This allows to make faster communications without a need to do expensive system calls for event loop wakeup.

Client (frontend) lifecycle

Whole client logic is driven by a single od_frontend() function, which is a coroutine entry point. There are 6 distinguishable stages in client lifecycle.

sources/frontend.h, sources/frontend.c

1. Startup

Read initial client request. This can be SSLRequest, CancelRequest or StartupMessage. Handle SSL/TLS handshake.

2. Process Cancel request

In case of CancelRequest, call Router to handle it. Disconnect client right away.

3. Route client

Call router. Use Database and User to match client configuration route. Router assigns matched route to a client. Each route object has a reference counter. All routes are periodically garbage-collected.

4. Authenticate client

Write client an authentication request AuthenticationMD5Password or AuthenticationCleartextPassword and wait for reply to compare passwords. In case of success send AuthenticationOk.

5. Process client requests

Depending on selected route storage type, do local (console) or remote (remote PostgreSQL server) processing.

Following remote processing logic repeats until client sends Terminate, client or server disconnects during the process:

  • Read client request. Handle Terminate.
  • If client has no server attached, call Router to assign server from the server pool. New server connection registered and initiated by the client coroutine (worker thread). Maybe discard previous server settings and configure it using client parameters.
  • Send client request to the server.
  • Wait for server reply.
  • Send reply to client.
  • In case of Transactional pooling: if transaction completes, call Router to detach server from the client.
  • Repeat.

6. Cleanup

If server is not Ready (query still in-progress), initiate automatic Cancel procedure. If server is Ready and left in active transaction, initiate automatic Rollback. Return server back to server pool or disconnect.

Free client context.

Client error codes

In the most scenarios PostgreSQL error messages ErrorResponce are copied to a client as-is. Yet, there are some cases, when Odyssey has to provide its own error message and SQLCode to client.

Function od_frontend_error() is used for formatting and sending error message to client.

SQLCode PostgreSQL Code Name Stage
08P01 PROTOCOL_VIOLATION Startup, TLS handshake, authentication
0A000 FEATURE_NOT_SUPPORTED TLS handshake
28000 INVALID_AUTHORIZATION_SPECIFICATION Authentication
28P01 INVALID_PASSWORD Authentication
58000 SYSTEM_ERROR Routing, System specific
3D000 UNDEFINED_DATABASE Routing
53300 TOO_MANY_CONNECTIONS Routing
08006 CONNECTION_FAILURE Server-side error during connection or IO

PostgreSQL specific error codes can be found in src/backend/errocodes.txt.