Megatest

;; (max-connections 4096)

(define (tt:get-conn ttdat dbfname)
  (hash-table-ref/default (tt-conns ttdat) dbfname #f))

;; do all the busy work of finding and setting up conn for
;; connecting to a server
;; This function, `tt:client-connect-to-server`, is designed to manage connections between a client and a server within a testing framework.
;; The function takes four arguments:
;; 1. `ttdat`: a data structure that holds information about the testing environment or connections.
;; 2. `dbfname`: The name of the database file that the client wants to connect to.
;; 3. `run-id`: An identifier for the current run of the test suite.
;; 4. `testsuite`: The test suite that is being run.
;; 
;;
;; Here's a step-by-step explanation of what the function does:
;;
;; 1. It first asserts that the `run-id` is valid for the given `dbfname` using the `tt:valid-run-id` function. If the `run-id` is not valid, it raises a fatal error.
;; 2. It prints debug information indicating that the function `tt:client-connect-to-server` has been called with the given `dbfname`.
;; 3. It attempts to retrieve an existing connection to the server from a hash table (`tt-conns`) using the `dbfname` as the key. If a connection already exists, it prints debug information and returns the existing connection.
;; 4. If no existing connection is found, it retrieves the current server information from the servinfo file, using the `tt:get-current-server-info` function.
;; 5. It uses pattern matching to destructure the server information into variables (`host`, `port`, `start-time`, `server-id`, `pid`, `dbfname2`, `servinffile`). It then asserts that the `dbfname` from the server info matches the one provided to the function.
;; 6. It constructs a connection object (`conn`) with the server information.
;; 7. It attempts to ping the server using `tt:timed-ping` to verify that the server is running and can be communicated with.
;; 8. Depending on the result of the ping:
;;    - If the server is running (`running`), it prints debug information, saves the connection in the hash table, and returns the connection.
;;    - If the server is starting (`starting`), it sleeps for 2 seconds and then recursively calls itself to retry the connection.
;;    - If the server is neither running nor starting, it checks if it's been more than 10 seconds since the last server start attempt. If so, it attempts to start the server using `server-start-proc` and then sleeps for 1 second before retrying the connection.
;; 9. If no server information is found (`else` case), it checks if it's been more than 3 seconds since the last server start attempt. If so, it starts a new server using `server-start-proc`, updates the last server start time, and sleeps for 4 seconds.
;; 10. It then sleeps for 1 second and prints debug information before recursively calling itself to retry the connection.
;;
;; The function uses recursion to keep trying to connect to the server, with various sleep intervals to prevent overwhelming the system with connection attempts or server starts.
;; It also uses a hash table to cache connections and avoid reconnecting to a server if a connection already exists.
;; The function is designed to handle different server states and ensure that a server is running and available before returning a valid connection to the caller.
;; 
(define (tt:client-connect-to-server ttdat dbfname run-id testsuite server-start-proc)
  (assert (tt:valid-run-id run-id dbfname) "FATAL: invalid run-id "run-id)
  (debug:print-info 2 *default-log-port* "tt:client-connect-to-server " dbfname " " run-id)
  (debug:print-info 2 *default-log-port* "tt:client-connect-to-server " dbfname)
  (let* ((conn              (tt:get-conn ttdat dbfname))
	 (server-start-proc (or server-start-proc
				(lambda ()
				  (assert (equal? dbfname "main.db") ;; only main.db is started here
					  "FATAL: called server-start-proc for db other than main.db")
				  (tt:server-process-run
				   (tt-areapath ttdat)
				   testsuite ;; (dbfile:testsuite-name)
				   (common:find-local-megatest)
				   run-id)))))
    (if conn
	(begin 
          (debug:print-info 2 *default-log-port* "already connected to a server")
          (debug:print-info 2 *default-log-port* "already connected to a server for " dbfname)
           conn) ;; we are already connected to the server

	;; no conn

	;; find server with lowest number of threads running (i.e. lowest load)
	;;
        (let* ((sdats (tt:get-server-info-sorted ttdat dbfname))
	       (sdat  (if (null? sdats)
			  #f
			  ;; choose server with lowest threads count
			  (car sdats))))
	   (debug:print-info 2 *default-log-port* "found sdat " sdat)
           (match sdat
			  (car (sort sdats
				     (lambda (a b)
				       (let* ((load-a (tt:get-server-threads a))
					      (load-b (tt:get-server-threads b)))
					 (< load-a load-b))))))))
				     
	  ;; (let ((indx (max (random (- (length sdats) 1)) 0)))
	  ;;    (list-ref sdats indx)))))
	  ;; (debug:print-info 1 *default-log-port* "found sdat " sdat" from sdats: "sdats)
          (match sdat
	    ((host port start-time server-id pid dbfname2 servinffile)
	     (assert (equal? dbfname dbfname2) "FATAL: read server info from wrong file.")
             (debug:print-info 2 *default-log-port* "no conn - in match servinffile:" servinffile)
	     (let* ((host-port (conc host":"port))
		    (conn (make-tt-conn
			   host: host
			   port: port

	     (tt:client-connect-to-server ttdat dbfname run-id testsuite server-start-proc)))))))

;; returns ( result . ping_time )
(define (tt:timed-ping host port server-id)
  (let* ((start-time (current-milliseconds))
	 (result     (tt:ping host port server-id)))
    (cons result (- (current-milliseconds) start-time))))
    

;; host:port => ( meta . when-updated)
(define *server-load* (make-hash-table))

(define (tt:save-server-meta host port meta)
  (hash-table-set! *server-load* (conc host":"port) (cons meta (current-seconds))))

(define (tt:get-server-threads dat)
  (let* ((host (car  dat))
	 (port (cadr dat))
	 (dat  (tt:get-server-meta host port #t)))
    ;; (debug:print 0 *default-log-port* "host: "host" port: "port" dat: "dat)
    (if (list? dat)
	(or (alist-ref 'sload dat) 99998)
	99999))) ;; absurd number means don't use this one

;; lazy get, does not auto-refresh meta, this might be a problem
;;
(define (tt:get-server-meta host port #!optional (do-ping #f))
  (let* ((get-meta (lambda ()
		     (let* ((dat  (hash-table-ref/default *server-load* (conc host":"port) #f)))
		       (if dat (car dat) #f))))
	 (meta     (get-meta)))
    (if (and (not meta)
	     do-ping)
	(begin
	  (tt:timed-ping host port #f)
	  (get-meta))
	meta)))

(define (tt:wait-on-server-load run-id ttdat)
  (if ttdat ;; if no server yet just pass on through
      (let* ((dbfname                 (dbmod:run-id->dbfname run-id))
	     (get-lowest-thread-load
	      (lambda ()
		(let* ((sdats (tt:get-server-info-sorted ttdat dbfname)))
		  (car (map tt:get-server-threads sdats))))))
	(if ttdat
	    (let loop ((count 0))
	      (let* ((lowestload (get-lowest-thread-load)))
		(if (> lowestload 5) ;; load is pretty high
		    (begin
		      (debug:print 0 *default-log-port* "Servers appear overloaded with "lowestload" threads, waiting...")
		      (thread-sleep! 1)
		      (if (< count 10)
			  (loop (+ count 1)))))))
	    (debug:print 0 *default-log-port* "Can't wait on server load, *ttdat* not set")))))

(define (tt:ping host port server-id #!optional (tries-left 5))
  (let*  ((res      (tt:send-receive-direct host port `(ping #f #f #f) ping-mode: #t)) ;; please send me your server-id
	  (try-again (lambda ()
		       (if (> tries-left 0)
			   (begin
			     (thread-sleep! 1)
			     (tt:ping host port server-id (- tries-left 1)))
			   #f))))
    ;;
    ;; need two threads, one a 5 second timer
    ;;
    (match res
      ((status errmsg result meta)
       (tt:save-server-meta host port meta)
       (if (equal? result server-id)
	   (let* ((server-state (alist-ref 'sstate meta)))
	     ;; (debug:print 0 *default-log-port* "Ping to "host":"port" successful.")
	     (or server-state 'unk)) ;; then we are good
	   (begin
	     (if server-id
	     (debug:print 0 *default-log-port* "WARNING: server-id does not match, expected: "server-id", got: "result)
		 (debug:print 0 *default-log-port* "WARNING: server-id does not match, expected: "server-id", got: "result))
	     #f)))
      (else
       ;; (debug:print 0 *default-log-port* "res not in form (status errmsg result meta), got: "res)
       (try-again)))))

;; client side handler
;;
;;(tt:handler #<tt> get-keys #f () 2 #f "/home/matt/data/megatest/ext-tests" #f "main.db" "ext-tests" "/home/matt/data/megatest/bin/.22.04/../megatest")
;;
;;g
(define (tt:handler ttdat cmd run-id params attemptnum readonly-mode dbfname testsuite mtexe server-start-proc)
  ;; connect-to-server will start a server if needed.
  (let* ((areapath (tt-areapath ttdat))
	 (conn     (tt:client-connect-to-server ttdat dbfname run-id testsuite server-start-proc))) ;; looks up conn keyed by dbfname
    (if conn
	;; have connection, call the server
	(let* ((res (tt:send-receive ttdat conn cmd run-id params)))

			 (debug:print 0 *default-log-port* "Server is loaded, delaying "delay-wait" seconds")
			 (thread-sleep! delay-wait)))))
	     (case status
	       ((busy) ;; result will be how long the server wants you to delay
		(let* ((raw-dly  (if (number? result) result 0.1))
		       (dly      (+ raw-dly (/ attemptnum 10)))) ;; (* raw-dly (/ attemptnum 2))))
		  (debug:print 0 *default-log-port* "WARNING: server for "dbfname" is busy, cmd is "cmd", will try again in "dly" seconds. This is attempt "(- attemptnum 1))
                  (debug:print 0 *default-log-port* errmsg)
		  (thread-sleep! dly)
		  (tt:handler  ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe server-start-proc)))
	       ((loaded)
		(debug:print 0 *default-log-port* "WARNING: server for "dbfname" is loaded, slowing queries.")
		(tt:backoff-incr (tt-conn-host conn)(tt-conn-port conn))

		;; this would be a good place to force reconnection and connect to a different server
		
		result) ;; (tt:handler  ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe))
	       (else
		result)))
	    (else ;; did not receive properly formated result
	     (if (not res) ;; tt:send-receive telling us that communication failed
		 (let* ((host    (tt-conn-host conn))
			(port    (tt-conn-port conn))

		   ;; (tt:handler ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe)
		   )))))
	(begin
	  (thread-sleep! 1) ;; no conn yet set up, give it a rest and try again
	  (tt:handler ttdat cmd run-id params attemptnum readonly-mode dbfname testsuite mtexe server-start-proc)))))

;; gets server info and appends path to server file
;; sorts by age, oldest first
;; sorts by age, --oldest-- now newest first
;;
;; move the ping here?
;;
;; returns list of (host port startseconds server-id servinfofile)
;;
(define (tt:get-server-info-sorted ttdat dbfname)
  (let* ((areapath (tt-areapath ttdat))
	 (sfiles   (tt:find-server areapath dbfname))
	 (sdats    (filter car (map tt:server-get-info sfiles))) ;; first element is #f if the file disappeared while being read
	 (sorted   (sort sdats (lambda (a b)
				 (let* ((starta (list-ref a 2))
					(startb (list-ref b 2)))
				   (if (eq? starta startb)
				       (string>? (list-ref a 3)(list-ref b 3)) ;; if servers started at same time look at server-id
				       (< starta startb))))))
				       (> starta startb))))))
	 (count    0))
    (for-each
     (lambda (rec)
       (if (or (> (length sorted) 1)
	       (common:low-noise-print 120 "server info sorted"))
	   (debug:print 2 *default-log-port* "SERVER #"count": "(string-intersperse (map conc sorted) ", ")))
       (set! count (+ count 1)))

;;======================================================================
;; server
;;======================================================================

(define (tt:sync-dbs ttdat)
  #f)

(define *server-start-requests* '())

;; start the listener and start responding to requests
;;
;; NOTE: organise by dbfname, not run-id so we don't need
;;       to pull in more modules
;;
;; This is the routine called in megatest.scm to start a server. NOTE: sequence is different for main.db vs. X.db

				     (same-host (or (not prime-host) ;; i.e. this is the first host
						    (equal? prime-host host)))
				     (keep-srv  (and good-ping same-host)))
				(if keep-srv	
				    (loop (cdr servrs)
					  host
					  (cons servdat result))
				    (begin
                                      ;; (debug:print-info 0 *default-log-port* "good-ping: " good-ping " same-host: " same-host "keep-srv: " keep-srv)
				      (handle-exceptions
				       exn
				       (debug:print-info 0 *default-log-port* "Error removing server info file: "servinfofile", "
							 (condition->list exn))
				       (delete-file* servinfofile))
				      (loop (cdr servrs) prime-host result)))))
				    (let* ((modtime (file-modification-time servinfofile)))
				      ;; if the .servinfo hasn't been touched in five min
				      ;; we can be pretty sure the server is truly dead
				      (if (> (- (current-seconds) modtime) 360)
					  (handle-exceptions
					   exn
					   (debug:print-info 0 *default-log-port*
							     "Error removing server info file: "servinfofile", "
							     (condition->list exn))
					   (delete-file* servinfofile))
					  (loop (cdr servrs) prime-host result))))))
			     (else
			      ;; can't delete it as we don't have a filename. NOTE: Should really never get here.
			      ;; can't delete it as we don't have a filename. NOTE: Should never get here.
			      (debug:print-info 0 *default-log-port* "ERROR: bad servinfo record \""servdat"\"")
			      (loop (cdr servrs) prime-host result)) ;; drop 
			     )))))
	       (home-host (if (null? good-srvrs)
			      #f
			      (caar good-srvrs))))
	  ;; by here we have a trustworthy list of servers and we have removed the .servinfo file for any unresponsive servers
	  ;; and the list is in good-srvrs
	  ;;
	  (cond
	   ((not home-host) ;; no servers yet, go ahead and start
	    (debug:print-info 0 *default-log-port* "No servers yet, starting on "(get-host-name)))
	   ((> (length good-srvrs) 2) ;; don't need more, just exit
	   ((> (length good-srvrs) 3) ;; don't need more, just exit
	    (debug:print-info 0 *default-log-port* "Have "(length good-srvrs)", no need for more, exiting.")
	    (exit))
	   ((not (equal? home-host (get-host-name))) ;; there is a home-host and we are not on it
	    (debug:print-info 0 *default-log-port* "Prime main server is on host "home-host", but we are on host "(get-host-name)", exiting.")
	    (exit))
	   (else
	    (debug:print-info 0 *default-log-port* "Starting on host "(get-host-name)", along with "(length good-srvrs)" other servers.")))

		      (thread-sleep! 0.25)
		      (loop (+ count 1)))
		    (begin
		      (debug:print 0 *default-log-port* "ERROR: (tt-port ttdat) no port set! Exiting.")
		      (exit)))))
	  
	  ;; create a servinfo file start keep-running
	  ;; On WSL there seems to be a race condition where the .servinfo file
	  ;; is not created fast enough
          (debug:print 0 *default-log-port* "Creating servinfo file for " dbfname)
	  (tt:create-server-registration-file ttdat dbfname)
	  (procinf-status-set! *procinf* "running")
	  (tt-state-set! ttdat 'running)
	  (dbfile:with-no-sync-db
	   nosyncdbpath
	   (lambda (nsdb)

	   nosyncdbpath
	   (lambda (nsdb)
	     (dbfile:insert-or-update-process nsdb *procinf*)))
	  (debug:print 0 *default-log-port* "Exiting now.")
	  (exit))))))

(define (tt:keep-running ttdat dbfname dbstruct)

  (thread-sleep! 1)
  
  ;; at this point the server is running and responding to calls, we just monitor
  ;; for db calls and exit if there are none.

  ;; if I am not in the first 3 servers, exit
  (let* ((start-time (current-seconds)))
    (let loop ()
      (let* ((servers   (tt:get-server-info-sorted ttdat dbfname))
	     (home-host (if (null? servers)
			    #f
			    (caar servers)))
	     (my-index  (list-index (lambda (x)
				      (equal? (list-ref x 6)
					      (tt-servinf-file ttdat)))
				    servers))
	     (ok         (cond
			  ((not my-index)
			   (debug:print 0 *default-log-port* "WARNING: Apparently I don't exist.")
			   #f) ;; keep trying or give up?
			  ((not *server-run*)
			   (debug:print 0 *default-log-port* "WARNING: received a stop server from client by remote request.")
			   #f)
			  ((null? servers)
			   (debug:print 0 *default-log-port* "WARNING: no servinfo files found, this cannot be.")
			   #f) ;; not ok
			  ((> my-index 2)
			   (debug:print 0 *default-log-port* "WARNING: there are more than two servers ahead of me, I'm not needed, exiting.")
			  ((> my-index 3)
			   (debug:print 0 *default-log-port* "WARNING: there are more than three servers ahead of me, I'm not needed, exiting.")
			   #f) ;; not ok to not be in first three
			  ((eq? (tt-state ttdat) 'running) #t) ;; we are good to keep going
			  ((> (- (current-seconds) start-time) 30)
			   (debug:print 0 *default-log-port* "WARNING: over 30 seconds and not yet in runnning mode. Exiting.")
			   #f)
			  (else #t))))
	(if ok

	      (let* ((sinfo-file (tt-servinf-file ttdat)))
		;; (debug:print 0 *default-log-port* "INFO: touching "sinfo-file)
		(set! (file-modification-time sinfo-file) (current-seconds))
		((dbr:dbstruct-sync-proc dbstruct) last-update)
		(dbr:dbstruct-last-update-set! dbstruct curr-secs))))
	
	(if (< (- (current-seconds) (tt-last-access ttdat)) (tt-server-timeout-param))
	    ;; process any requests to start a new server due to load on this one
	    (let* ((requests *server-start-requests*))
	      (set! *server-start-requests* '())
	      (if (> (length requests) 0)
		  (debug:print-info 0 *default-log-port* "Processing "(length requests)" server start requests"))
	      (for-each (lambda (proc)
	    (begin
			  (proc)
			  (thread-sleep! 1))
			requests)
	      (thread-sleep! 5)
	      (loop)))))
    (tt:shutdown-server ttdat)
    (debug:print 0 *default-log-port* "INFO: Server timed out, exiting from tt:keep-running.")))


(define (tt:shutdown-server ttdat)

	 (servinf (conc servdir"/"host":"port"-"(current-process-id)":"dbfname))
	 (serv-id (tt:mk-signature areapath)))
    (assert (and host port) "FATAL: tt:create-server-registration-file called with no conn, dbfname="dbfname)
    (tt-servinf-file-set! ttdat servinf)
    (with-output-to-file servinf
      (lambda ()
	(print "SERVER STARTED: "host":"port" AT "(current-seconds)" server-id: "serv-id" pid: "(current-process-id)" dbfname: "dbfname)))
    (let loop ((count 0))
      (if (not (file-exists? servinf))
	  (begin
	    (debug:print 0 *default-log-port* "WARNING: file "servinf" was created but it doesn't show up on disk! We'll try again.")
	    (thread-sleep! 1)
	    (if (> count 10)
		(debug:print 0 *default-log-port* "WARNING: file "servinf" was not created.")
		(loop (+ count 1))))))
      serv-id))
    serv-id))

;; find valid server
;; get servers listed, last part of name must match :<dbfname>
;; if more than one, wait one second and look again
;; 
;; NOTE: this only gets the servinfo data, no network activity here
;; future: ping oldest, if alive remove other :<dbfname> files
;;       i.e. no ping etc.
;;
(define (tt:find-server areapath dbfname)
  (let* ((servdir  (tt:get-servinfo-dir areapath))
	 (sfiles   (glob (conc servdir"/*:"dbfname)))
	 (goodfiles '()))

    ;; filter the files here by looking in processes table (if we are not main.db)

(define *last-server-start* (make-hash-table))

(define (tt:too-recent-server-start dbfname)
  (let* ((last-run-time (hash-table-ref/default *last-server-start* dbfname #f)))
    (and last-run-time
	 (< (- (current-seconds) last-run-time) 5))))

(define *last-server-start-request-time* 0)
    
;; Given an area path,  start a server process    ### NOTE ### > file 2>&1 
;; if the target-host is set 
;; try running on that host
;;   incidental: rotate logs in logs/ dir.
;;
(define  (tt:server-process-run areapath testsuite mtexe run-id #!key (profile-mode "")) ;; areapath is *toppath* for a given testsuite area
  (assert areapath  "FATAL: tt:server-process-run called without areapath defined.")
  (assert testsuite "FATAL: tt:server-process-run called without testsuite defined.")
  (assert mtexe     "FATAL: tt:server-process-run called without mtexe defined.")
  ;; mtest -server - -m testsuite:ext-tests -db 6.db
  (let* ((dbfname  (dbmod:run-id->dbfname run-id)))
    (if (or (< (- (current-seconds) *last-server-start-request-time*) 5) ;; attempted start less than 5 sec ago
    (if (tt:too-recent-server-start dbfname)
	    (tt:too-recent-server-start dbfname))
	#f
	(let* ((load     (get-normalized-cpu-load))
	       (srvrs    (tt:find-server areapath dbfname))
	       (trying   (length srvrs))
	       (nrun     (number-of-processes-running (conc "mtest.*server.*"testsuite".*"dbfname))))
	  (set! *last-server-start-request-time* (current-seconds))
	  (cond
	   ((> load 2.0)
	    (debug:print 0 *default-log-port* "Normalized load "load" on " (get-host-name) " is over the limit of 2.0. Not starting a server. Please reduce the load on "(get-host-name)" by killing some processes")
	    (thread-sleep! 1)
	   ((> load 4.0)
	    (debug:print 0 *default-log-port* "Normalized load " load " over 4, (load: " (commonmod:get-cpu-load) " cores: " (get-current-host-cores) " exiting...")
	    (thread-sleep! 1) ;; I'm not convinced that a delay here is helpful. -mrw-
	    #f)
	   ((> nrun 100)
	    (debug:print 0 *default-log-port* nrun" servers running on " (get-host-name) ", not starting another.")
	    (thread-sleep! 1)
	    #f)
	   ((> trying 2)
	   ((> trying 3)
	    (debug:print 0 *default-log-port* trying" servers registered in .servinfo dir. not starting another.")
	    (thread-sleep! 1)
	    #f)
	   (else
	    (if (not (file-exists? (conc areapath"/logs")))
		(create-directory (conc areapath"/logs") #t))
	    (let* ((logfile   (conc areapath "/logs/server-"dbfname"-"(current-process-id)".log")) ;; -" curr-pid "-" target-host ".log"))