if too many errors on a task, stop trying

README.md

@@ -148,11 +148,11 @@ contributions welcome,,,,
 ## status
 
 ### base
-- 🚧 input space manipulation functions
+- ✅ input space manipulation functions
     - ✅ data types: using data/integer-set, pattern (vector of integer-set)
     - ✅ basic manipulation functions
     - ✅ representation of input space as a flat integer
-- 🚧 #lang for configuration/definitions
+- ✅ #lang for configuration/definitions
     - (input) mode
         - stdio: user program gets input by stdio, integers separated by space, one per line
         - callback: input generator compiled into user program, user main calls `crossfire_main`
@@ -161,8 +161,8 @@ contributions welcome,,,,
     - SMP: performed by crossfire or performed by the user code
         - "performed by user code" can also mean GPU, for example
 - 🚧 codegen for input generator (in C)
-    - ✅ stdio mode
-    - 🚧 callback mode
+    - 🚧 stdio mode
+    - ✅ callback mode
     - ✅ success reporting mechanism
     - low priority: configurable "character" type -- currently a "character" is a uint64\_t
 
@@ -172,7 +172,7 @@ contributions welcome,,,,
 - low priority: randomized input space distribution
 - low priority: store common configuration templates for clients
 - low priority: track upload/download progress
-- streaming interface for file transfers
+- ✅ streaming interface for file transfers
 - ✅ accept submitted projects (with client-compiled input generator) and distribute to agents
     - ✅ low priority: support for multiple architectures
 - ✅ agent authentication

crossfire/server.rkt

@@ -407,6 +407,11 @@
 (define *max-retry-delay* 120)
 (define *retry-delay-ratio* 2)
 
+;; every time agent reports an error, we increment error count
+;; every time there is success, we decrement the error count (to 0)
+;; threshold for error count before we stop trying to run this task
+(define *max-task-errors* 10)
+
 (define (agent-handler)
   ;; unlike comms, messages to agent-handler have no responses. just thread-send, it's one-way
   (define cust (make-custodian))
@@ -427,7 +432,7 @@
   ;; precisely than just the taskid, which allows them to cache the file locally
   ;; completed-work: an integer set of completed work
   (struct task-state [id sema manifest [work-pattern #:mutable] agent-todo file-hash
-                      [completed-work #:mutable]] #:transparent)
+                      [completed-work #:mutable] error-log] #:transparent)
 
   (define (initialize-task id mf)
     (define file-hash (server-hash-file id))
@@ -441,7 +446,8 @@
                 ([(_1 _2 _3 pat-fasl) (in-query (current-db) q-get-task-log id)])
         (define sub (make-integer-set (fasl->s-exp pat-fasl)))
         (values (integer-set-subtract pattern-range sub) (integer-set-union completed-work sub))))
-    (task-state id sema mf pattern-range agent-todo file-hash completed-work))
+    (define error-log (make-hash))
+    (task-state id sema mf pattern-range agent-todo file-hash completed-work error-log))
 
   (define (task-has-work? ts)
     (not (and (zero? (integer-set-count (task-state-work-pattern ts)))
@@ -498,8 +504,10 @@
           assignment]))))
 
   ;; returns work from agent back to the regular work pool
-  (define (task-unassign! ts agent-id)
+  (define (task-unassign! ts agent-id [error? #f])
     (call-with-semaphore (task-state-sema ts) (lambda ()
+      (when error?
+        (hash-update! (task-state-error-log ts) agent-id add1 0))
       (match (hash-ref (task-state-agent-todo ts) agent-id #f)
         [#f (void)]
         [assignment
@@ -511,6 +519,7 @@
   ;; adds to task log, then updates work pool with task completion
   (define (task-complete! ts agent-id time-wall-start duration)
     (call-with-semaphore (task-state-sema ts) (lambda ()
+      (hash-update! (task-state-error-log ts) agent-id (lambda (x) (max 0 (sub1 x))) 0)
       (match (hash-ref (task-state-agent-todo ts) agent-id #f)
         [#f (void)]
         [assignment
@@ -619,12 +628,12 @@
                             (integer-set-contents assign-data)))
          #t]))
 
-    (define (cancel-assignment! assignment)
+    (define (cancel-assignment! assignment [error? #f])
       ;; tell the agent to cancel work, unassign the assignment
       (define ts (hash-ref current-tasks (assignment-taskid assignment) #f))
       ;; update manager tracking
       (unless (false? ts)
-        (task-unassign! ts id))
+        (task-unassign! ts id error?))
       (hash-remove! assigned-tasks (assignment-id assignment))
       ;; send agent rpc
       (with-handlers ([exn:fail?
@@ -705,10 +714,12 @@
           ['() #f]
           [(cons head tail)
            (define manifest (task-state-manifest head))
+           (define error-count (hash-ref (task-state-error-log head) id 0))
            (define needed-resources (list->set (manifest-data-ref manifest 'resources '())))
            (define needed-arch (manifest-data-ref manifest 'arch '("any")))
            (define right-arch? (or (member "any" needed-arch) (member arch needed-arch)))
            (if (and right-arch?
+                    (< error-count *max-task-errors*)
                     (task-has-work? head)
                     (subset? needed-resources available-resources))
              (create-assignment! head)
@@ -747,11 +758,12 @@
             (unless (false? av) (complete-assignment! av))
             (update-assignments!)]
            ;; execution failed, unassign task
-           ;; TODO : maybe blacklist this task from this agent if there are too many errors
            ;; TODO : notify connected clients that an error occurred
            [(cons assignment-id 'error)
             (define av (hash-ref assigned-tasks assignment-id #f))
-            (unless (false? av) (cancel-assignment! av))
+            (unless (false? av)
+              ;; #t for yes, this is an error
+              (cancel-assignment! av #t))
             (update-assignments!)]
            ;; got succeeding input
            [(cons assignment-id success-input)