Simplify vectorization task to sequential processing

Drop the threaded worker pool (queue, mutexes, connection pooling).
The task now processes packages one by one via find_each - plenty
fast since the external embedding service dominates latency, and
much easier to reason about.
This commit is contained in:
Christoph Haas 2026-08-23 00:13:50 +02:00
parent 3550a4e2e2
commit 2b79470feb

View file

@ -5,21 +5,12 @@
# rubocop:disable Metrics/BlockLength
namespace :debshots do
desc 'Compute embedding vectors for all packages based on their description ' \
'and long description (missing ones only unless FORCE=1). ' \
'Set CONCURRENCY=n to control parallel HTTP requests.'
'and long description (missing ones only unless FORCE=1)'
task compute_vectors: :environment do
logger = Logger.new($stdout)
logger.level = Logger::INFO
force = ENV['FORCE'].present?
concurrency = [ENV.fetch('CONCURRENCY', '4').to_i, 1].max
# Never use more threads than the connection pool has connections.
# Keep one connection reserved for the main thread.
pool_size = ActiveRecord::Base.connection_pool.size
concurrency = [concurrency, pool_size - 1].min
concurrency = 0 if concurrency.negative?
scope = force ? Package.all : Package.where(embedding: nil)
total = scope.count
@ -30,61 +21,33 @@ namespace :debshots do
logger.info "Computing vectors for #{total} packages " \
"(#{force ? 'recomputing' : 'missing only'}, " \
"service: #{Vectorizer.service_url}, " \
"concurrency: #{concurrency.zero? ? 1 : concurrency})"
"service: #{Vectorizer.service_url})"
stats_mutex = Mutex.new
stats = { done: 0, failed: 0 }
done = 0
failed = 0
process_package = lambda do |package|
scope.find_each do |package|
begin
raise ArgumentError, 'no description text' if package.embedding_text.blank?
package.update_embedding!
stats_mutex.synchronize { stats[:done] += 1 }
done += 1
rescue StandardError => e
# A single broken package must never take down a whole worker
# thread - log enough detail to diagnose it and move on.
stats_mutex.synchronize { stats[:failed] += 1 }
# A single broken package must never stop the whole run - log
# enough detail to diagnose it and move on.
failed += 1
logger.error "Failed for package #{package.name}: #{e.class}: #{e.message}"
logger.error e.backtrace.first(3).join("\n") if e.backtrace
end
stats_mutex.synchronize do
processed = stats[:done] + stats[:failed]
return unless (processed % 100).zero?
processed = done + failed
next unless (processed % 100).zero? || processed == total
percent = (processed.to_f / total * 100).round(1)
logger.info "Progress: #{processed}/#{total} (#{percent}%)"
end
percent = (processed.to_f / total * 100).round(1)
logger.info "Progress: #{processed}/#{total} (#{percent}%)"
end
if concurrency.zero?
# Sequential fallback when the connection pool has no room for workers
scope.find_each { |package| process_package.call(package) }
else
# Feed package IDs into a queue that the worker threads pull from.
work_queue = Queue.new
scope.pluck(:id).each { |id| work_queue << id }
concurrency.times { work_queue << nil } # one poison pill per worker
workers = Array.new(concurrency) do
Thread.new do
# Each thread gets its own database connection from the pool.
ActiveRecord::Base.connection_pool.with_connection do
while (package_id = work_queue.pop)
package = Package.find_by(id: package_id)
process_package.call(package) if package
end
end
rescue StandardError => e
logger.error "Worker crashed: #{e.message}"
end
end
workers.each(&:join)
end
logger.info "Done. #{stats[:done]} vectors computed, #{stats[:failed]} failures."
logger.info "Done. #{done} vectors computed, #{failed} failures."
end
end
# rubocop:enable Metrics/BlockLength