kemingy

Data Types

https://redis.io/topics/data-types-intro

• binary-safe strings
• lists: strings (insertion order) (linked list instead of array)
• sets: strings
• sorted sets: (strings, score)
• hashes: {string:string}
• bit array: bitmaps

kemingy

Vagrant.configure("2") do |config|

  config.vm.box = "hashicorp/bionic64"

  config.vm.define :pd do |pd|
    pd.vm.hostname = "pingcap.pd"
    pd.vm.network :private_network, ip: "192.168.0.1"
    pd.vm.network "forwarded_port", guest: 2379, host: 2379, host_ip: "127.0.0.1"
    pd.vm.network "forwarded_port", guest: 3000, host: 3000, host_ip: "127.0.0.1"
    pd.vm.network "forwarded_port", guest: 3000, host: 9093, host_ip: "127.0.0.1"

kemingy

import logging
import signal
import subprocess
import sys
from pathlib import Path

logger = logging.getLogger(__name__)


def run(filepath):

kemingy

Training

• learning rate
• dropout
• max token
• clip norm
• tokenization method
• update freq (mini-batch with delayed update)
• optimizer
• learning rate scheduler

kemingy

import socket
import numpy as np
import pickle
import time
import struct

matrix = np.random.random((224, 224, 3))
socket_path = './uds.sock'

kemingy

• from image version FROM python:3.8.5-buster
• package version fixed in requirements.txt file
• build cache
• non-root user

  RUN useradd --create-home appuser
  WORKDIR /home/appuser
  USER appuser
  

• don't use alpine (slow, bug)

kemingy

import multiprocessing
from time import time
from functools import wraps


def timeit(epoch=1000):
    def decorator(func):
        @wraps(func)
        def wrap(*args, **kw):
            ts = time()

kemingy

x = 233


def xxx():
    x = 222

    def foo():
        nonlocal x
        x = 1

kemingy

• Lock
• RLock: a reentrant lock
• Condition:
  • wait: release the lock
  • notify/notify_all
• Semaphore: manages an internal counter (always non-negative)
• Event: manages an internal flag (boolean)
• Timer
• Barrier: a fixed number of threads that need to wait for each other

kemingy

Features Required

• prevent single point faliure
• batch requests
• scaling
• job recovery (retry)

Queue