queue.lua

-------------------------------------------------------------------------------
-- Queue class
-------------------------------------------------------------------------------
-- Return a queue object
function Qless.queue(name)
  assert(name, 'Queue(): no queue name provided')
  local queue = {}
  setmetatable(queue, QlessQueue)
  queue.name = name

  -- Access to our work
  queue.work = {
    peek = function(now, offset, count)
      if count == 0 then
        return {}
      end
      local jids = {}
      for _, jid in ipairs(redis.call('zrevrange', queue:prefix('work'), offset, offset + count - 1)) do
        table.insert(jids, jid)
      end
      return jids
    end,
    remove = function(...)
      if select('#', ...) > 0 then
        return redis.call('zrem', queue:prefix('work'), ...)
      end
    end,
    add = function(now, priority, jid)
      if priority ~= '+inf' then
        priority = priority - (now / 10000000000)
      end
      return redis.call('zadd', queue:prefix('work'), priority, jid)
    end,
    score = function(jid)
      return redis.call('zscore', queue:prefix('work'), jid)
    end,
    length = function()
      return redis.call('zcard', queue:prefix('work'))
    end,
  }

  -- Access to our locks
  queue.locks = {
    expired = function(now, offset, count)
      return redis.call('zrangebyscore', queue:prefix('locks'), '-inf', now, 'LIMIT', offset, count)
    end,
    peek = function(now, offset, count)
      return redis.call('zrangebyscore', queue:prefix('locks'), now, '+inf', 'LIMIT', offset, count)
    end,
    add = function(expires, jid)
      redis.call('zadd', queue:prefix('locks'), expires, jid)
    end,
    remove = function(...)
      if select('#', ...) > 0 then
        return redis.call('zrem', queue:prefix('locks'), ...)
      end
    end,
    running = function(now)
      return redis.call('zcount', queue:prefix('locks'), now, '+inf')
    end,
    length = function(now)
      -- If a 'now' is provided, we're interested in how many are before
      -- that time
      if now then
        return redis.call('zcount', queue:prefix('locks'), 0, now)
      else
        return redis.call('zcard', queue:prefix('locks'))
      end
    end,
    job_time_left = function(now, jid)
      return tonumber(redis.call('zscore', queue:prefix('locks'), jid) or 0) - now
    end,
  }

  -- Access to our dependent jobs
  queue.depends = {
    peek = function(now, offset, count)
      return redis.call('zrange', queue:prefix('depends'), offset, offset + count - 1)
    end,
    add = function(now, jid)
      redis.call('zadd', queue:prefix('depends'), now, jid)
    end,
    remove = function(...)
      if select('#', ...) > 0 then
        return redis.call('zrem', queue:prefix('depends'), ...)
      end
    end,
    length = function()
      return redis.call('zcard', queue:prefix('depends'))
    end,
  }

  -- Access to our scheduled jobs
  queue.scheduled = {
    peek = function(now, offset, count)
      return redis.call('zrange', queue:prefix('scheduled'), offset, offset + count - 1)
    end,
    ready = function(now, offset, count)
      return redis.call('zrangebyscore', queue:prefix('scheduled'), 0, now, 'LIMIT', offset, count)
    end,
    add = function(when, jid)
      redis.call('zadd', queue:prefix('scheduled'), when, jid)
    end,
    remove = function(...)
      if select('#', ...) > 0 then
        return redis.call('zrem', queue:prefix('scheduled'), ...)
      end
    end,
    length = function()
      return redis.call('zcard', queue:prefix('scheduled'))
    end,
  }

  -- Access to our recurring jobs
  queue.recurring = {
    peek = function(now, offset, count)
      return redis.call('zrangebyscore', queue:prefix('recur'), 0, now, 'LIMIT', offset, count)
    end,
    add = function(when, jid)
      redis.call('zadd', queue:prefix('recur'), when, jid)
    end,
    remove = function(...)
      if select('#', ...) > 0 then
        return redis.call('zrem', queue:prefix('recur'), ...)
      end
    end,
    update = function(increment, jid)
      redis.call('zincrby', queue:prefix('recur'), increment, jid)
    end,
    score = function(jid)
      return redis.call('zscore', queue:prefix('recur'), jid)
    end,
    length = function()
      return redis.call('zcard', queue:prefix('recur'))
    end,
  }
  return queue
end

-- Return the prefix for this particular queue
function QlessQueue:prefix(group)
  return QlessQueue.ns .. self.name .. (group and ('-' .. group) or '')
end

-- Stats(now, date)
-- ---------------------
-- Return the current statistics for a given queue on a given date. The
-- results are returned are a JSON blob:
--
--
--  {
--      'failed': 3,
--      'retries': 5,
--      'wait' : {
--          'total'    : ...,
--          'mean'     : ...,
--          'variance' : ...,
--          'histogram': [
--              ...
--          ]
--      }, 'run': {
--          'total'    : ...,
--          'mean'     : ...,
--          'variance' : ...,
--          'histogram': [
--              ...
--          ]
--      }
--  }
--
-- The histogram's data points are at the second resolution for the first
-- minute, the minute resolution for the first hour, the 15-minute resolution
-- for the first day, the hour resolution for the first 3 days, and then at
-- the day resolution from there on out. The `histogram` key is a list of
-- those values.
function QlessQueue:stats(now, date)
  date = assert(tonumber(date), 'Stats(): Arg "date" missing or not a number: ' .. (date or 'nil'))

  -- The bin is midnight of the provided day
  -- 24 * 60 * 60 = 86400
  local bin = date - (date % 86400)

  -- This a table of all the keys we want to use in order to produce a histogram
  local histokeys = {}
  for _, unit in ipairs({ { 's', 0, 59 }, { 'm', 1, 59 }, { 'h', 1, 23 }, { 'd', 1, 6 } }) do
    for i = unit[2], unit[3] do
      table.insert(histokeys, unit[1] .. i)
    end
  end

  local mkstats = function(name, queue)
    -- The results we'll be sending back
    local results = {}

    local key = 'ql:s:' .. name .. ':' .. bin .. ':' .. queue
    local count, mean, vk = unpack(redis.call('hmget', key, 'total', 'mean', 'vk'))

    count = tonumber(count) or 0
    mean = tonumber(mean) or 0
    vk = tonumber(vk)

    results.count = count or 0
    results.mean = mean or 0
    results.histogram = {}

    if not count then
      results.std = 0
    else
      if count > 1 then
        results.std = math.sqrt(vk / (count - 1))
      else
        results.std = 0
      end
    end

    local histogram = redis.call('hmget', key, unpack(histokeys))
    for i = 1, #histokeys do
      table.insert(results.histogram, tonumber(histogram[i]) or 0)
    end
    return results
  end

  local retries, failed, failures =
    unpack(redis.call('hmget', 'ql:s:stats:' .. bin .. ':' .. self.name, 'retries', 'failed', 'failures'))
  return {
    retries = tonumber(retries or 0),
    failed = tonumber(failed or 0),
    failures = tonumber(failures or 0),
    wait = mkstats('wait', self.name),
    run = mkstats('run', self.name),
  }
end

-- Peek
-------
-- Examine the next jobs that would be popped from the queue without actually
-- popping them.
function QlessQueue:peek(now, count)
  count = assert(tonumber(count), 'Peek(): Arg "count" missing or not a number: ' .. tostring(count))

  -- These are the ids that we're going to return. We'll begin with any jobs
  -- that have lost their locks
  local jids = self.locks.expired(now, 0, count)

  -- If we still need jobs in order to meet demand, then we should
  -- look for all the recurring jobs that need jobs run
  self:check_recurring(now, count - #jids)

  -- Now we've checked __all__ the locks for this queue the could
  -- have expired, and are no more than the number requested. If
  -- we still need values in order to meet the demand, then we
  -- should check if any scheduled items, and if so, we should
  -- insert them to ensure correctness when pulling off the next
  -- unit of work.
  self:check_scheduled(now, count - #jids)

  -- With these in place, we can expand this list of jids based on the work
  -- queue itself and the priorities therein
  extend_table(jids, self.work.peek(now, 0, count - #jids))

  return jids
end

-- Return true if this queue is paused
function QlessQueue:paused()
  return redis.call('sismember', 'ql:paused_queues', self.name) == 1
end

-- Pause this queue
--
-- Note: long term, we have discussed adding a rate-limiting
-- feature to qless-core, which would be more flexible and
-- could be used for pausing (i.e. pause = set the rate to 0).
-- For now, this is far simpler, but we should rewrite this
-- in terms of the rate limiting feature if/when that is added.
function QlessQueue.pause(now, ...)
  redis.call('sadd', 'ql:paused_queues', ...)
end

-- Unpause this queue
function QlessQueue.unpause(...)
  redis.call('srem', 'ql:paused_queues', ...)
end

-- Checks for expired locks, scheduled and recurring jobs, returning any
-- jobs that are ready to be processes
function QlessQueue:pop(now, worker, count)
  assert(worker, 'Pop(): Arg "worker" missing')
  count = assert(tonumber(count), 'Pop(): Arg "count" missing or not a number: ' .. tostring(count))

  -- We should find the heartbeat interval for this queue heartbeat
  local expires = now + tonumber(Qless.config.get(self.name .. '-heartbeat') or Qless.config.get('heartbeat', 60))

  -- If this queue is paused, then return no jobs
  if self:paused() then
    return {}
  end

  -- Make sure we add this worker to the list of seen workers
  redis.call('zadd', 'ql:workers', now, worker)

  -- Check our max concurrency, and limit the count
  local max_concurrency = tonumber(Qless.config.get(self.name .. '-max-concurrency', 0))

  if max_concurrency > 0 then
    -- Allow at most max_concurrency - #running
    local allowed = math.max(0, max_concurrency - self.locks.running(now))
    count = math.min(allowed, count)
    if count == 0 then
      return {}
    end
  end

  local jids = self:invalidate_locks(now, count)
  -- Now we've checked __all__ the locks for this queue the could
  -- have expired, and are no more than the number requested.

  -- If we still need jobs in order to meet demand, then we should
  -- look for all the recurring jobs that need jobs run
  self:check_recurring(now, count - #jids)

  -- If we still need values in order to meet the demand, then we
  -- should check if any scheduled items, and if so, we should
  -- insert them to ensure correctness when pulling off the next
  -- unit of work.
  self:check_scheduled(now, count - #jids)

  -- With these in place, we can expand this list of jids based on the work
  -- queue itself and the priorities therein
  extend_table(jids, self.work.peek(now, 0, count - #jids))

  for _, jid in ipairs(jids) do
    local job = Qless.job(jid)
    job:history(now, 'popped', { worker = worker })

    -- Update the wait time statistics
    local time = tonumber(redis.call('hget', QlessJob.ns .. jid, 'time') or now)
    local waiting = now - time
    self:stat(now, 'wait', waiting)
    redis.call('hset', QlessJob.ns .. jid, 'time', string.format('%.20f', now))

    -- Add this job to the list of jobs handled by this worker
    redis.call('zadd', 'ql:w:' .. worker .. ':jobs', expires, jid)

    -- Update the jobs data, and add its locks, and return the job
    job:update({
      worker = worker,
      expires = expires,
      state = 'running',
    })

    self.locks.add(expires, jid)

    local tracked = redis.call('zscore', 'ql:tracked', jid) ~= false
    if tracked then
      Qless.publish('popped', jid)
    end
  end

  -- If we are returning any jobs, then we should remove them from the work
  -- queue
  self.work.remove(unpack(jids))

  return jids
end

-- Update the stats for this queue
function QlessQueue:stat(now, stat, val)
  -- The bin is midnight of the provided day
  local bin = now - (now % 86400)
  local key = 'ql:s:' .. stat .. ':' .. bin .. ':' .. self.name

  -- Get the current data
  local count, mean, vk = unpack(redis.call('hmget', key, 'total', 'mean', 'vk'))

  -- If there isn't any data there presently, then we must initialize it
  count = count or 0
  if count == 0 then
    mean = val
    vk = 0
    count = 1
  else
    count = count + 1
    local oldmean = mean
    mean = mean + (val - mean) / count
    vk = vk + (val - mean) * (val - oldmean)
  end

  -- Now, update the histogram
  -- - `s1`, `s2`, ..., -- second-resolution histogram counts
  -- - `m1`, `m2`, ..., -- minute-resolution
  -- - `h1`, `h2`, ..., -- hour-resolution
  -- - `d1`, `d2`, ..., -- day-resolution
  val = math.floor(val)
  if val < 60 then -- seconds
    redis.call('hincrby', key, 's' .. val, 1)
  elseif val < 3600 then -- minutes
    redis.call('hincrby', key, 'm' .. math.floor(val / 60), 1)
  elseif val < 86400 then -- hours
    redis.call('hincrby', key, 'h' .. math.floor(val / 3600), 1)
  else -- days
    redis.call('hincrby', key, 'd' .. math.floor(val / 86400), 1)
  end
  redis.call('hmset', key, 'total', count, 'mean', mean, 'vk', vk)
end

-- Put(now, jid, klass, data, delay,
--     [priority, p],
--     [tags, t],
--     [retries, r],
--     [depends, '[...]'])
-- -----------------------
-- Insert a job into the queue with the given priority, tags, delay, klass and
-- data.
function QlessQueue:put(now, worker, jid, klass, data, delay, ...)
  assert(jid, 'Put(): Arg "jid" missing')
  assert(klass, 'Put(): Arg "klass" missing')
  assert(cjson.decode(data), 'Put(): Arg "data" missing or not JSON: ' .. tostring(data))
  delay = assert(tonumber(delay), 'Put(): Arg "delay" not a number: ' .. tostring(delay))

  -- Read in all the optional parameters. All of these must come in pairs, so
  -- if we have an odd number of extra args, raise an error
  local arg_count = select('#', ...)
  if arg_count % 2 == 1 then
    error('Odd number of additional args')
  end
  local options = {}
  for i = 1, arg_count, 2 do
    local key, val = select(i, ...)
    options[key] = val
  end

  -- Let's see what the old priority and tags were
  local job = Qless.job(jid)
  local priority, tags, oldqueue, state, failure, retries, oldworker, interval, next_run, old_resources = unpack(
    redis.call(
      'hmget',
      QlessJob.ns .. jid,
      'priority',
      'tags',
      'queue',
      'state',
      'failure',
      'retries',
      'worker',
      'throttle_interval',
      'throttle_next_run',
      'resources'
    )
  )

  next_run = next_run or now

  -- true if empty or anything other than false
  local replace =
    assert(tonumber(options['replace'] or 1), 'Put(): Arg "replace" not a number: ' .. tostring(options['replace']))

  if replace == 0 and state == 'running' then
    local time_left = self.locks.job_time_left(now, jid)
    if time_left > 0 then
      return time_left
    end
  end

  -- If there are old tags, then we should remove the tags this job has
  if tags then
    Qless.tag(now, 'remove', jid, unpack(cjson.decode(tags)))
  end

  -- Sanity check on optional args
  retries = assert(
    tonumber(options['retries'] or retries or 5),
    'Put(): Arg "retries" not a number: ' .. tostring(options['retries'])
  )
  tags =
    assert(cjson.decode(options['tags'] or tags or '[]'), 'Put(): Arg "tags" not JSON' .. tostring(options['tags']))
  priority = assert(
    tonumber(options['priority'] or priority or 0),
    'Put(): Arg "priority" not a number' .. tostring(options['priority'])
  )
  local depends =
    assert(cjson.decode(options['depends'] or '[]'), 'Put(): Arg "depends" not JSON: ' .. tostring(options['depends']))

  local resources = assert(
    cjson.decode(options['resources'] or '[]'),
    'Put(): Arg "resources" not JSON array: ' .. tostring(options['resources'])
  )
  assert(#resources == 0 or QlessResource.all_exist(resources), 'Put(): invalid resources requested')

  -- if there were previously acquired resources, verify consistency
  if old_resources then
    old_resources = Set.new(cjson.decode(old_resources))
    local removed_resources = Set.diff(old_resources, Set.new(resources))
    for k in pairs(removed_resources) do
      Qless.resource(k):release(now, jid)
    end
  end

  interval = assert(
    tonumber(options['interval'] or interval or 0),
    'Put(): Arg "interval" not a number: ' .. tostring(options['interval'])
  )

  if interval > 0 then
    local minimum_delay = next_run - now
    if minimum_delay >= 0 then
      delay = math.max(delay, minimum_delay)
    end
  else
    next_run = 0
  end

  -- If the job has old dependencies, determine which dependencies are
  -- in the new dependencies but not in the old ones, and which are in the
  -- old ones but not in the new
  if #depends > 0 then
    -- This makes it easier to check if it's in the new list
    local new = {}
    for _, d in ipairs(depends) do
      new[d] = 1
    end

    -- Now find what's in the original, but not the new
    local original = redis.call('smembers', QlessJob.ns .. jid .. '-dependencies')
    for _, dep in pairs(original) do
      if not new[dep] then
        -- Remove k as a dependency
        redis.call('srem', QlessJob.ns .. dep .. '-dependents', jid)
        redis.call('srem', QlessJob.ns .. jid .. '-dependencies', dep)
      end
    end
  end

  -- Send out a log message
  Qless.publish(
    'log',
    cjson.encode({
      jid = jid,
      event = 'put',
      queue = self.name,
    })
  )

  -- Update the history to include this new change
  job:history(now, 'put', { q = self.name })

  -- If this item was previously in another queue, then we should remove it from there
  if oldqueue then
    local queue_obj = Qless.queue(oldqueue)
    queue_obj.work.remove(jid)
    queue_obj.locks.remove(jid)
    queue_obj.depends.remove(jid)
    queue_obj.scheduled.remove(jid)
  end

  -- If this had previously been given out to a worker, make sure to remove it
  -- from that worker's jobs
  if oldworker and oldworker ~= '' then
    redis.call('zrem', 'ql:w:' .. oldworker .. ':jobs', jid)
    -- If it's a different worker that's putting this job, send a notification
    -- to the last owner of the job
    if oldworker ~= worker then
      -- We need to inform whatever worker had that job
      local encoded = cjson.encode({
        jid = jid,
        event = 'lock_lost',
        worker = oldworker,
      })
      Qless.publish('w:' .. oldworker, encoded)
      Qless.publish('log', encoded)
    end
  end

  -- If the job was previously in the 'completed' state, then we should
  -- remove it from being enqueued for destruction
  if state == 'complete' then
    redis.call('zrem', 'ql:completed', jid)
  end

  -- Add this job to the list of jobs tagged with whatever tags were supplied
  for _, tag in ipairs(tags) do
    redis.call('zadd', 'ql:t:' .. tag, now, jid)
    redis.call('zincrby', 'ql:tags', 1, tag)
  end

  -- If we're in the failed state, remove all of our data
  if state == 'failed' then
    failure = cjson.decode(failure)
    -- We need to make this remove it from the failed queues
    redis.call('lrem', 'ql:f:' .. failure.group, 0, jid)
    if redis.call('llen', 'ql:f:' .. failure.group) == 0 then
      redis.call('srem', 'ql:failures', failure.group)
    end
    -- The bin is midnight of the provided day
    -- 24 * 60 * 60 = 86400
    local bin = failure.when - (failure.when % 86400)
    -- We also need to decrement the stats about the queue on
    -- the day that this failure actually happened.
    redis.call('hincrby', 'ql:s:stats:' .. bin .. ':' .. self.name, 'failed', -1)
  end

  -- First, let's save its data
  hmset(QlessJob.ns .. jid, {
    jid = jid,
    klass = klass,
    data = data,
    priority = priority,
    tags = cjson.encode(tags),
    resources = cjson.encode(resources),
    state = ((delay > 0) and 'scheduled') or 'waiting',
    worker = '',
    expires = 0,
    queue = self.name,
    retries = retries,
    remaining = retries,
    time = string.format('%.20f', now),
    throttle_interval = interval,
    throttle_next_run = next_run,
    result_data = '{}',
  })

  -- These are the jids we legitimately have to wait on
  for _, j in ipairs(depends) do
    -- Make sure it's something other than 'nil' or complete.
    local dep_state = redis.call('hget', QlessJob.ns .. j, 'state')
    if dep_state and dep_state ~= 'complete' then
      redis.call('sadd', QlessJob.ns .. j .. '-dependents', jid)
      redis.call('sadd', QlessJob.ns .. jid .. '-dependencies', j)
    end
  end

  -- Now, if a delay was provided, and if it's in the future,
  -- then we'll have to schedule it. Otherwise, we're just
  -- going to add it to the work queue.
  if delay > 0 then
    if redis.call('scard', QlessJob.ns .. jid .. '-dependencies') > 0 then
      -- We've already put it in 'depends'. Now, we must just save the data
      -- for when it's scheduled
      self.depends.add(now, jid)
      redis.call('hmset', QlessJob.ns .. jid, 'state', 'depends', 'scheduled', now + delay)
    else
      self.scheduled.add(now + delay, jid)
    end
  else
    if redis.call('scard', QlessJob.ns .. jid .. '-dependencies') > 0 then
      self.depends.add(now, jid)
      redis.call('hset', QlessJob.ns .. jid, 'state', 'depends')
    elseif #resources > 0 then
      if Qless.job(jid):acquire_resources(now) then
        self.work.add(now, priority, jid)
      end
    else
      self.work.add(now, priority, jid)
    end
  end

  -- Lastly, we're going to make sure that this item is in the
  -- set of known queues. We should keep this sorted by the
  -- order in which we saw each of these queues
  if redis.call('zscore', 'ql:queues', self.name) == false then
    redis.call('zadd', 'ql:queues', now, self.name)
  end

  if redis.call('zscore', 'ql:tracked', jid) ~= false then
    Qless.publish('put', jid)
  end

  return jid
end

-- Move `count` jobs out of the failed state and into this queue
function QlessQueue:unfail(now, group, count)
  assert(group, 'Unfail(): Arg "group" missing')
  count = assert(tonumber(count or 25), 'Unfail(): Arg "count" not a number: ' .. tostring(count))

  -- Get up to that many jobs, and we'll put them in the appropriate queue
  local jids = redis.call('lrange', 'ql:f:' .. group, -count, -1)

  -- And now set each job's state, and put it into the appropriate queue
  for _, jid in ipairs(jids) do
    local job = Qless.job(jid)
    local data = job:data()
    job:history(now, 'put', { q = self.name })
    hmset(QlessJob.ns .. data.jid, {
      state = 'waiting',
      worker = '',
      expires = 0,
      queue = self.name,
      remaining = data.retries or 5,
    })

    if job:acquire_resources(now) then
      self.work.add(now, data.priority, data.jid)
    end
  end

  -- Remove these jobs from the failed state
  redis.call('ltrim', 'ql:f:' .. group, 0, -count - 1)
  if redis.call('llen', 'ql:f:' .. group) == 0 then
    redis.call('srem', 'ql:failures', group)
  end

  return #jids
end

-- Recur a job of type klass in this queue
function QlessQueue:recur(now, jid, klass, data, spec, ...)
  assert(jid, 'RecurringJob On(): Arg "jid" missing')
  assert(klass, 'RecurringJob On(): Arg "klass" missing')
  assert(spec, 'RecurringJob On(): Arg "spec" missing')
  assert(cjson.decode(data), 'RecurringJob On(): Arg "data" missing or not JSON: ' .. tostring(data))

  -- At some point in the future, we may have different types of recurring
  -- jobs, but for the time being, we only have 'interval'-type jobs
  if spec == 'interval' then
    local interval, offset = ...
    interval = assert(tonumber(interval), 'Recur(): Arg "interval" not a number: ' .. tostring(interval))
    offset = assert(tonumber(offset), 'Recur(): Arg "offset" not a number: ' .. tostring(offset))
    if interval <= 0 then
      error('Recur(): Arg "interval" must be greater than 0')
    end

    -- Read in all the optional parameters. All of these must come in
    -- pairs, so if we have an odd number of extra args, raise an error
    local arg_count = select('#', ...)
    if arg_count % 2 == 1 then
      error('Odd number of additional args')
    end
    local options = {}
    for i = 3, arg_count, 2 do
      local key, val = select(i, ...)
      options[key] = val
    end
    options.tags = assert(
      cjson.decode(options.tags or '{}'),
      'Recur(): Arg "tags" must be JSON string array: ' .. tostring(options.tags)
    )
    options.priority =
      assert(tonumber(options.priority or 0), 'Recur(): Arg "priority" not a number: ' .. tostring(options.priority))
    options.retries =
      assert(tonumber(options.retries or 0), 'Recur(): Arg "retries" not a number: ' .. tostring(options.retries))
    options.backlog =
      assert(tonumber(options.backlog or 0), 'Recur(): Arg "backlog" not a number: ' .. tostring(options.backlog))
    options.resources = assert(
      cjson.decode(options['resources'] or '[]'),
      'Recur(): Arg "resources" not JSON array: ' .. tostring(options['resources'])
    )

    local count, old_queue = unpack(redis.call('hmget', 'ql:r:' .. jid, 'count', 'queue'))
    count = count or 0

    -- If it has previously been in another queue, then we should remove
    -- some information about it
    if old_queue then
      Qless.queue(old_queue).recurring.remove(jid)
    end

    -- Do some insertions
    hmset('ql:r:' .. jid, {
      jid = jid,
      klass = klass,
      data = data,
      priority = options.priority,
      tags = cjson.encode(options.tags or {}),
      state = 'recur',
      queue = self.name,
      type = 'interval',
      -- How many jobs we've spawned from this
      count = count,
      interval = interval,
      retries = options.retries,
      backlog = options.backlog,
      resources = cjson.encode(options.resources),
    })
    -- Now, we should schedule the next run of the job
    self.recurring.add(now + offset, jid)

    -- Lastly, we're going to make sure that this item is in the
    -- set of known queues. We should keep this sorted by the
    -- order in which we saw each of these queues
    if redis.call('zscore', 'ql:queues', self.name) == false then
      redis.call('zadd', 'ql:queues', now, self.name)
    end

    return jid
  else
    error('Recur(): schedule type "' .. tostring(spec) .. '" unknown')
  end
end

-- Return the length of the queue
function QlessQueue:length()
  return self.locks.length() + self.work.length() + self.scheduled.length()
end

-------------------------------------------------------------------------------
-- Housekeeping methods
-------------------------------------------------------------------------------
-- Instantiate any recurring jobs that are ready
function QlessQueue:check_recurring(now, count)
  -- This is how many jobs we've moved so far
  local moved = 0
  -- These are the recurring jobs that need work
  local r = self.recurring.peek(now, 0, count)
  for _, jid in ipairs(r) do
    -- For each of the jids that need jobs scheduled, first
    -- get the last time each of them was run, and then increment
    -- it by its interval. While this time is less than now,
    -- we need to keep putting jobs on the queue
    local klass, data, priority, tags, retries, interval, backlog, resources = unpack(
      redis.call(
        'hmget',
        'ql:r:' .. jid,
        'klass',
        'data',
        'priority',
        'tags',
        'retries',
        'interval',
        'backlog',
        'resources'
      )
    )
    local _tags = cjson.decode(tags)
    resources = cjson.decode(resources or '[]')
    local score = math.floor(tonumber(self.recurring.score(jid)) or 0)
    interval = tonumber(interval)

    -- If the backlog is set for this job, then see if it's been a long
    -- time since the last pop
    backlog = tonumber(backlog or 0)
    if backlog ~= 0 then
      -- Check how many jobs we could conceivably generate
      local num = ((now - score) / interval)
      if num > backlog then
        -- Update the score
        score = score + (math.ceil(num - backlog) * interval)
      end
    end

    -- We're saving this value so that in the history, we can accurately
    -- reflect when the job would normally have been scheduled
    while (score <= now) and (moved < count) do
      local spawn_count = redis.call('hincrby', 'ql:r:' .. jid, 'count', 1)
      moved = moved + 1

      local child_jid = jid .. '-' .. spawn_count

      -- Add this job to the list of jobs tagged with whatever tags were
      -- supplied
      for _, tag in ipairs(_tags) do
        redis.call('zadd', 'ql:t:' .. tag, now, child_jid)
        redis.call('zincrby', 'ql:tags', 1, tag)
      end

      -- First, let's save its data
      hmset(QlessJob.ns .. child_jid, {
        jid = child_jid,
        klass = klass,
        data = data,
        priority = priority,
        tags = tags,
        state = 'waiting',
        worker = '',
        expires = 0,
        queue = self.name,
        retries = retries,
        remaining = retries,
        resources = cjson.encode(resources),
        throttle_interval = 0,
        time = string.format('%.20f', score),
        spawned_from_jid = jid,
      })

      local job = Qless.job(child_jid)
      job:history(score, 'put', { q = self.name })

      -- Now, if a delay was provided, and if it's in the future,
      -- then we'll have to schedule it. Otherwise, we're just
      -- going to add it to the work queue.

      local add_job = true
      if #resources then
        add_job = job:acquire_resources(score)
      end
      if add_job then
        self.work.add(score, priority, child_jid)
      end

      score = score + interval
      self.recurring.add(score, jid)
    end
  end
end

-- Check for any jobs that have been scheduled, and shovel them onto
-- the work queue. Returns nothing, but afterwards, up to `count`
-- scheduled jobs will be moved into the work queue
function QlessQueue:check_scheduled(now, count)
  -- zadd is a list of arguments that we'll be able to use to
  -- insert into the work queue
  local scheduled = self.scheduled.ready(now, 0, count)
  for _, jid in ipairs(scheduled) do
    -- With these in hand, we'll have to go out and find the
    -- priorities of these jobs, and then we'll insert them
    -- into the work queue and then when that's complete, we'll
    -- remove them from the scheduled queue
    local priority = tonumber(redis.call('hget', QlessJob.ns .. jid, 'priority') or 0)
    if Qless.job(jid):acquire_resources(now) then
      self.work.add(now, priority, jid)
    end
    self.scheduled.remove(jid)

    -- We should also update them to have the state 'waiting'
    -- instead of 'scheduled'
    redis.call('hset', QlessJob.ns .. jid, 'state', 'waiting')
  end
end

-- Check for and invalidate any locks that have been lost. Returns the
-- list of jids that have been invalidated
function QlessQueue:invalidate_locks(now, count)
  local jids = {}
  -- Iterate through all the expired locks and add them to the list
  -- of keys that we'll return
  for _, jid in ipairs(self.locks.expired(now, 0, count)) do
    -- Remove this job from the jobs that the worker that was running it
    -- has
    local worker = unpack(redis.call('hmget', QlessJob.ns .. jid, 'worker'))
    redis.call('zrem', 'ql:w:' .. worker .. ':jobs', jid)

    -- We'll provide a grace period after jobs time out for them to give
    -- some indication of the failure mode. After that time, however, we'll
    -- consider the worker dust in the wind
    local grace_period = tonumber(Qless.config.get('grace-period'))

    -- Whether or not we've already sent a courtesy message
    local courtesy_sent = tonumber(redis.call('hget', QlessJob.ns .. jid, 'grace') or 0)

    -- If the remaining value is an odd multiple of 0.5, then we'll assume
    -- that we're just sending the message. Otherwise, it's time to
    -- actually hand out the work to another worker
    local send_message = (courtesy_sent ~= 1)
    local invalidate = not send_message

    -- If the grace period has been disabled, then we'll do both.
    if grace_period <= 0 then
      send_message = true
      invalidate = true
    end

    if send_message then
      -- This is where we supply a courtesy message and give the worker
      -- time to provide a failure message
      if redis.call('zscore', 'ql:tracked', jid) ~= false then
        Qless.publish('stalled', jid)
      end
      Qless.job(jid):history(now, 'timed-out')
      redis.call('hset', QlessJob.ns .. jid, 'grace', 1)

      -- Send a message to let the worker know that it's lost its lock on
      -- the job
      local encoded = cjson.encode({
        jid = jid,
        event = 'lock_lost',
        worker = worker,
      })
      Qless.publish('w:' .. worker, encoded)
      Qless.publish('log', encoded)
      self.locks.add(now + grace_period, jid)

      -- If we got any expired locks, then we should increment the
      -- number of retries for this stage for this bin. The bin is
      -- midnight of the provided day
      local bin = now - (now % 86400)
      redis.call('hincrby', 'ql:s:stats:' .. bin .. ':' .. self.name, 'retries', 1)
    end

    if invalidate then
      -- Unset the grace period attribute so that next time we'll send
      -- the grace period
      redis.call('hdel', QlessJob.ns .. jid, 'grace', 0)

      -- See how many remaining retries the job has
      local remaining = tonumber(redis.call('hincrby', QlessJob.ns .. jid, 'remaining', -1))

      -- This is where we actually have to time out the work
      if remaining < 0 then
        Qless.job(jid):release_resources(now)

        -- Now remove the instance from the schedule, and work queues
        -- for the queue it's in
        self.work.remove(jid)
        self.locks.remove(jid)
        self.scheduled.remove(jid)

        local group = 'failed-retries-' .. self.name
        local job = Qless.job(jid)
        job:history(now, 'failed', { group = group })
        redis.call('hmset', QlessJob.ns .. jid, 'state', 'failed', 'worker', '', 'expires', '')
        -- If the failure has not already been set, then set it
        redis.call(
          'hset',
          QlessJob.ns .. jid,
          'failure',
          cjson.encode({
            ['group'] = group,
            ['message'] = 'Job exhausted retries in queue "' .. self.name .. '"',
            ['when'] = now,
            ['worker'] = worker,
          })
        )

        -- Add this type of failure to the list of failures
        redis.call('sadd', 'ql:failures', group)
        -- And add this particular instance to the failed types
        redis.call('lpush', 'ql:f:' .. group, jid)

        if redis.call('zscore', 'ql:tracked', jid) ~= false then
          Qless.publish('failed', jid)
        end
        Qless.publish(
          'log',
          cjson.encode({
            jid = jid,
            event = 'failed',
            group = group,
            worker = worker,
            message = 'Job exhausted retries in queue "' .. self.name .. '"',
          })
        )

        -- Increment the count of the failed jobs
        local bin = now - (now % 86400)
        redis.call('hincrby', 'ql:s:stats:' .. bin .. ':' .. self.name, 'failures', 1)
        redis.call('hincrby', 'ql:s:stats:' .. bin .. ':' .. self.name, 'failed', 1)
      else
        table.insert(jids, jid)
      end
    end
  end

  return jids
end

-- Forget the provided queues. As in, remove them from the list of known queues
function QlessQueue.deregister(...)
  redis.call('zrem', Qless.ns .. 'queues', ...)
end

-- Return information about a particular queue, or all queues
--  [
--      {
--          'name': 'testing',
--          'stalled': 2,
--          'waiting': 5,
--          'running': 5,
--          'scheduled': 10,
--          'depends': 5,
--          'recurring': 0
--      }, {
--          ...
--      }
--  ]
function QlessQueue.counts(now, name)
  if name then
    local queue = Qless.queue(name)
    local stalled = queue.locks.length(now)
    -- Check for any scheduled jobs that need to be moved
    queue:check_scheduled(now, queue.scheduled.length())
    return {
      name = name,
      waiting = queue.work.length(),
      stalled = stalled,
      running = queue.locks.length() - stalled,
      scheduled = queue.scheduled.length(),
      depends = queue.depends.length(),
      recurring = queue.recurring.length(),
      paused = queue:paused(),
    }
  else
    local queues = redis.call('zrange', 'ql:queues', 0, -1)
    local response = {}
    for _, qname in ipairs(queues) do
      table.insert(response, QlessQueue.counts(now, qname))
    end
    return response
  end
end