# # OtterTune - fabfile.py # # Copyright (c) 2017-18, Carnegie Mellon University Database Group # ''' Created on Mar 23, 2018 @author: bohan ''' import glob import json import os import re import time from collections import OrderedDict from multiprocessing import Process import logging from logging.handlers import RotatingFileHandler import requests from fabric.api import env, lcd, local, settings, show, task from fabric.state import output as fabric_output from utils import (file_exists, get, load_driver_conf, parse_bool, put, run, run_sql_script, sudo, FabricException) # Loads the driver config file (defaults to driver_config.py) dconf = load_driver_conf() # pylint: disable=invalid-name # Fabric settings fabric_output.update({ 'running': True, 'stdout': True, }) env.abort_exception = FabricException env.hosts = [dconf.LOGIN] # Create local directories for _d in (dconf.RESULT_DIR, dconf.LOG_DIR, dconf.TEMP_DIR): os.makedirs(_d, exist_ok=True) # Configure logging LOG = logging.getLogger(__name__) LOG.setLevel(getattr(logging, dconf.LOG_LEVEL, logging.DEBUG)) Formatter = logging.Formatter( # pylint: disable=invalid-name fmt='%(asctime)s [%(funcName)s:%(lineno)03d] %(levelname)-5s: %(message)s', datefmt='%m-%d-%Y %H:%M:%S') ConsoleHandler = logging.StreamHandler() # pylint: disable=invalid-name ConsoleHandler.setFormatter(Formatter) LOG.addHandler(ConsoleHandler) FileHandler = RotatingFileHandler( # pylint: disable=invalid-name dconf.DRIVER_LOG, maxBytes=50000, backupCount=2) FileHandler.setFormatter(Formatter) LOG.addHandler(FileHandler) @task def check_disk_usage(): partition = dconf.DATABASE_DISK disk_use = 0 if partition: cmd = "df -h {}".format(partition) out = run(cmd).splitlines()[1] m = re.search(r'\d+(?=%)', out) if m: disk_use = int(m.group(0)) LOG.info("Current Disk Usage: %s%s", disk_use, '%') return disk_use @task def check_memory_usage(): run('free -m -h') @task def create_controller_config(): if dconf.DB_TYPE == 'postgres': dburl_fmt = 'jdbc:postgresql://{host}:{port}/{db}'.format elif dconf.DB_TYPE == 'oracle': dburl_fmt = 'jdbc:oracle:thin:@{host}:{port}:{db}'.format else: raise Exception("Database Type {} Not Implemented !".format(dconf.DB_TYPE)) config = dict( database_type=dconf.DB_TYPE, database_url=dburl_fmt(host=dconf.DB_HOST, port=dconf.DB_PORT, db=dconf.DB_NAME), username=dconf.DB_USER, password=dconf.DB_PASSWORD, upload_code='DEPRECATED', upload_url='DEPRECATED', workload_name=dconf.OLTPBENCH_BENCH ) with open(dconf.CONTROLLER_CONFIG, 'w') as f: json.dump(config, f, indent=2) @task def restart_database(): if dconf.DB_TYPE == 'postgres': if dconf.HOST_CONN == 'docker': # Restarting the docker container here is the cleanest way to do it # becaues there's no init system running and the only process running # in the container is postgres itself local('docker restart {}'.format(dconf.CONTAINER_NAME)) else: sudo('pg_ctl -D {} -w -t 600 restart -m fast'.format( dconf.PG_DATADIR), user=dconf.ADMIN_USER, capture=False) elif dconf.DB_TYPE == 'oracle': run_sql_script('restartOracle.sh') else: raise Exception("Database Type {} Not Implemented !".format(dconf.DB_TYPE)) @task def drop_database(): if dconf.DB_TYPE == 'postgres': run("PGPASSWORD={} dropdb -e --if-exists {} -U {} -h {}".format( dconf.DB_PASSWORD, dconf.DB_NAME, dconf.DB_USER, dconf.DB_HOST)) else: raise Exception("Database Type {} Not Implemented !".format(dconf.DB_TYPE)) @task def create_database(): if dconf.DB_TYPE == 'postgres': run("PGPASSWORD={} createdb -e {} -U {} -h {}".format( dconf.DB_PASSWORD, dconf.DB_NAME, dconf.DB_USER, dconf.DB_HOST)) else: raise Exception("Database Type {} Not Implemented !".format(dconf.DB_TYPE)) @task def create_user(): if dconf.DB_TYPE == 'postgres': sql = "CREATE USER {} SUPERUSER PASSWORD '{}';".format(dconf.DB_USER, dconf.DB_PASSWORD) run("PGPASSWORD={} psql -c \\\"{}\\\" -U postgres -h {}".format( dconf.DB_PASSWORD, sql, dconf.DB_HOST)) elif dconf.DB_TYPE == 'oracle': run_sql_script('createUser.sh', dconf.DB_USER, dconf.DB_PASSWORD) else: raise Exception("Database Type {} Not Implemented !".format(dconf.DB_TYPE)) @task def drop_user(): if dconf.DB_TYPE == 'postgres': sql = "DROP USER IF EXISTS {};".format(dconf.DB_USER) run("PGPASSWORD={} psql -c \\\"{}\\\" -U postgres -h {}".format( dconf.DB_PASSWORD, sql, dconf.DB_HOST)) elif dconf.DB_TYPE == 'oracle': run_sql_script('dropUser.sh', dconf.DB_USER) else: raise Exception("Database Type {} Not Implemented !".format(dconf.DB_TYPE)) @task def reset_conf(): change_conf() @task def change_conf(next_conf=None): signal = "# configurations recommended by ottertune:\n" next_conf = next_conf or {} tmp_conf_in = os.path.join(dconf.TEMP_DIR, os.path.basename(dconf.DB_CONF) + '.in') get(dconf.DB_CONF, tmp_conf_in) with open(tmp_conf_in, 'r') as f: lines = f.readlines() if signal not in lines: lines += ['\n', signal] signal_idx = lines.index(signal) lines = lines[0:signal_idx + 1] if dconf.BASE_DB_CONF: assert isinstance(dconf.BASE_DB_CONF, dict), \ (type(dconf.BASE_DB_CONF), dconf.BASE_DB_CONF) base_conf = ['{} = {}\n'.format(*c) for c in sorted(dconf.BASE_DB_CONF.items())] lines.extend(base_conf) if isinstance(next_conf, str): with open(next_conf, 'r') as f: recommendation = json.load( f, encoding="UTF-8", object_pairs_hook=OrderedDict)['recommendation'] else: recommendation = next_conf assert isinstance(recommendation, dict) for name, value in recommendation.items(): if dconf.DB_TYPE == 'oracle' and isinstance(value, str): value = value.strip('B') lines.append('{} = {}\n'.format(name, value)) lines.append('\n') tmp_conf_out = os.path.join(dconf.TEMP_DIR, os.path.basename(dconf.DB_CONF) + '.out') with open(tmp_conf_out, 'w') as f: f.write(''.join(lines)) sudo('cp {0} {0}.ottertune.bak'.format(dconf.DB_CONF)) put(tmp_conf_out, dconf.DB_CONF, use_sudo=False) local('rm -f {} {}'.format(tmp_conf_in, tmp_conf_out)) @task def load_oltpbench(): cmd = "./oltpbenchmark -b {} -c {} --create=true --load=true".\ format(dconf.OLTPBENCH_BENCH, dconf.OLTPBENCH_CONFIG) with lcd(dconf.OLTPBENCH_HOME): # pylint: disable=not-context-manager local(cmd) @task def run_oltpbench(): cmd = "./oltpbenchmark -b {} -c {} --execute=true -s 5 -o outputfile".\ format(dconf.OLTPBENCH_BENCH, dconf.OLTPBENCH_CONFIG) with lcd(dconf.OLTPBENCH_HOME): # pylint: disable=not-context-manager local(cmd) @task def run_oltpbench_bg(): cmd = "./oltpbenchmark -b {} -c {} --execute=true -s 5 -o outputfile > {} 2>&1 &".\ format(dconf.OLTPBENCH_BENCH, dconf.OLTPBENCH_CONFIG, dconf.OLTPBENCH_LOG) with lcd(dconf.OLTPBENCH_HOME): # pylint: disable=not-context-manager local(cmd) @task def run_controller(): if not os.path.exists(dconf.CONTROLLER_CONFIG): create_controller_config() cmd = 'gradle run -PappArgs="-c {} -d output/" --no-daemon > {}'.\ format(dconf.CONTROLLER_CONFIG, dconf.CONTROLLER_LOG) with lcd(dconf.CONTROLLER_HOME): # pylint: disable=not-context-manager local(cmd) @task def signal_controller(): pidfile = os.path.join(dconf.CONTROLLER_HOME, 'pid.txt') with open(pidfile, 'r') as f: pid = int(f.read()) cmd = 'kill -2 {}'.format(pid) with lcd(dconf.CONTROLLER_HOME): # pylint: disable=not-context-manager local(cmd) @task def save_dbms_result(): t = int(time.time()) files = ['knobs.json', 'metrics_after.json', 'metrics_before.json', 'summary.json'] for f_ in files: srcfile = os.path.join(dconf.CONTROLLER_HOME, 'output', f_) dstfile = os.path.join(dconf.RESULT_DIR, '{}__{}'.format(t, f_)) local('cp {} {}'.format(srcfile, dstfile)) return t @task def save_next_config(next_config, t=None): if not t: t = int(time.time()) with open(os.path.join(dconf.RESULT_DIR, '{}__next_config.json'.format(t)), 'w') as f: json.dump(next_config, f, indent=2) return t @task def free_cache(): if dconf.HOST_CONN != 'docker': with show('everything'), settings(warn_only=True): # pylint: disable=not-context-manager res = sudo("sh -c \"echo 3 > /proc/sys/vm/drop_caches\"") if res.failed: LOG.error('%s (return code %s)', res.stderr.strip(), res.return_code) @task def upload_result(result_dir=None, prefix=None, upload_code=None): result_dir = result_dir or os.path.join(dconf.CONTROLLER_HOME, 'output') prefix = prefix or '' upload_code = upload_code or dconf.UPLOAD_CODE files = {} for base in ('summary', 'knobs', 'metrics_before', 'metrics_after'): fpath = os.path.join(result_dir, prefix + base + '.json') # Replaces the true db version with the specified version to allow for # testing versions not officially supported by OtterTune if base == 'summary' and dconf.OVERRIDE_DB_VERSION: with open(fpath, 'r') as f: summary = json.load(f) summary['real_database_version'] = summary['database_version'] summary['database_version'] = dconf.OVERRIDE_DB_VERSION with open(fpath, 'w') as f: json.dump(summary, f, indent=1) files[base] = open(fpath, 'rb') response = requests.post(dconf.WEBSITE_URL + '/new_result/', files=files, data={'upload_code': upload_code}) if response.status_code != 200: raise Exception('Error uploading result.\nStatus: {}\nMessage: {}\n'.format( response.status_code, response.content)) for f in files.values(): # pylint: disable=not-an-iterable f.close() LOG.info(response.content) return response @task def get_result(max_time_sec=180, interval_sec=5, upload_code=None): max_time_sec = int(max_time_sec) interval_sec = int(interval_sec) upload_code = upload_code or dconf.UPLOAD_CODE url = dconf.WEBSITE_URL + '/query_and_get/' + upload_code elapsed = 0 response_dict = None response = '' while elapsed <= max_time_sec: rsp = requests.get(url) response = rsp.content.decode() assert response != 'null' LOG.debug('%s [status code: %d, content_type: %s, elapsed: %ds]', response, rsp.status_code, rsp.headers.get('content-type', ''), elapsed) if rsp.status_code == 200: # Success response_dict = json.loads(rsp.json(), object_pairs_hook=OrderedDict) break elif rsp.status_code == 202: # Not ready time.sleep(interval_sec) elapsed += interval_sec elif rsp.status_code == 400: # Failure raise Exception( "Failed to download the next config.\nStatus code: {}\nMessage: {}\n".format( rsp.status_code, response)) else: raise NotImplementedError( "Unhandled status code: '{}'.\nMessage: {}".format(rsp.status_code, response)) if not response_dict: assert elapsed > max_time_sec, \ 'response={} but elapsed={}s <= max_time={}s'.format( response, elapsed, max_time_sec) raise Exception( 'Failed to download the next config in {}s: {} (elapsed: {}s)'.format( max_time_sec, response, elapsed)) LOG.info('Downloaded the next config in %ds: %s', elapsed, json.dumps(response_dict, indent=4)) return response_dict @task def download_debug_info(pprint=False): pprint = parse_bool(pprint) url = '{}/dump/{}'.format(dconf.WEBSITE_URL, dconf.UPLOAD_CODE) params = {'pp': int(True)} if pprint else {} rsp = requests.get(url, params=params) if rsp.status_code != 200: raise Exception('Error downloading debug info.') filename = rsp.headers.get('Content-Disposition').split('=')[-1] file_len, exp_len = len(rsp.content), int(rsp.headers.get('Content-Length')) assert file_len == exp_len, 'File {}: content length != expected length: {} != {}'.format( filename, file_len, exp_len) with open(filename, 'wb') as f: f.write(rsp.content) LOG.info('Downloaded debug info to %s', filename) return filename @task def add_udf(): local('python3 ./LatencyUDF.py ../controller/output/') @task def upload_batch(result_dir=None, sort=True, upload_code=None): result_dir = result_dir or dconf.RESULT_DIR sort = parse_bool(sort) results = glob.glob(os.path.join(result_dir, '*__summary.json')) if sort: results = sorted(results) count = len(results) LOG.info('Uploading %d samples from %s...', count, result_dir) for i, result in enumerate(results): prefix = os.path.basename(result) prefix_len = os.path.basename(result).find('_') + 2 prefix = prefix[:prefix_len] upload_result(result_dir=result_dir, prefix=prefix, upload_code=upload_code) LOG.info('Uploaded result %d/%d: %s__*.json', i + 1, count, prefix) @task def dump_database(): dumpfile = os.path.join(dconf.DB_DUMP_DIR, dconf.DB_NAME + '.dump') if file_exists(dumpfile): LOG.info('%s already exists ! ', dumpfile) return False LOG.info('Dump database %s to %s', dconf.DB_NAME, dumpfile) if dconf.DB_TYPE == 'oracle': run_sql_script('dumpOracle.sh', dconf.DB_USER, dconf.DB_PASSWORD, dconf.DB_NAME, dconf.DB_DUMP_DIR) elif dconf.DB_TYPE == 'postgres': run('PGPASSWORD={} pg_dump -U {} -h {} -F c -d {} > {}'.format( dconf.DB_PASSWORD, dconf.DB_USER, dconf.DB_HOST, dconf.DB_NAME, dumpfile)) else: raise Exception("Database Type {} Not Implemented !".format(dconf.DB_TYPE)) return True @task def restore_database(): dumpfile = os.path.join(dconf.DB_DUMP_DIR, dconf.DB_NAME + '.dump') if not file_exists(dumpfile): raise FileNotFoundError("Database dumpfile '{}' does not exist!".format(dumpfile)) LOG.info('Start restoring database') if dconf.DB_TYPE == 'oracle': # You must create a directory named dpdata through sqlplus in your Oracle database # The following script assumes such directory exists. drop_user() create_user() run_sql_script('restoreOracle.sh', dconf.DB_USER, dconf.DB_NAME) elif dconf.DB_TYPE == 'postgres': drop_database() create_database() run('PGPASSWORD={} pg_restore -U {} -h {} -n public -j 8 -F c -d {} {}'.format( dconf.DB_PASSWORD, dconf.DB_USER, dconf.DB_HOST, dconf.DB_NAME, dumpfile)) else: raise Exception("Database Type {} Not Implemented !".format(dconf.DB_TYPE)) LOG.info('Finish restoring database') def _ready_to_start_oltpbench(): ready = False if os.path.exists(dconf.CONTROLLER_LOG): with open(dconf.CONTROLLER_LOG, 'r') as f: content = f.read() ready = 'Output the process pid to' in content return ready def _ready_to_start_controller(): ready = False if os.path.exists(dconf.OLTPBENCH_LOG): with open(dconf.OLTPBENCH_LOG, 'r') as f: content = f.read() ready = 'Warmup complete, starting measurements' in content return ready def _ready_to_shut_down_controller(): pidfile = os.path.join(dconf.CONTROLLER_HOME, 'pid.txt') ready = False if os.path.exists(pidfile) and os.path.exists(dconf.OLTPBENCH_LOG): with open(dconf.OLTPBENCH_LOG, 'r') as f: content = f.read() ready = 'Output throughput samples into file' in content return ready def clean_logs(): # remove oltpbench and controller log files local('rm -f {} {}'.format(dconf.OLTPBENCH_LOG, dconf.CONTROLLER_LOG)) @task def loop(i): i = int(i) # free cache free_cache() # remove oltpbench log and controller log clean_logs() # restart database restart_database() time.sleep(dconf.RESTART_SLEEP_SEC) # check disk usage if check_disk_usage() > dconf.MAX_DISK_USAGE: LOG.warning('Exceeds max disk usage %s', dconf.MAX_DISK_USAGE) # run controller from another process p = Process(target=run_controller, args=()) p.start() LOG.info('Run the controller') # run oltpbench as a background job while not _ready_to_start_oltpbench(): time.sleep(1) run_oltpbench_bg() LOG.info('Run OLTP-Bench') # the controller starts the first collection while not _ready_to_start_controller(): time.sleep(1) signal_controller() LOG.info('Start the first collection') # stop the experiment while not _ready_to_shut_down_controller(): time.sleep(1) signal_controller() LOG.info('Start the second collection, shut down the controller') p.join() # add user defined target objective # add_udf() # save result result_timestamp = save_dbms_result() if i >= dconf.WARMUP_ITERATIONS: # upload result upload_result() # get result response = get_result() # save next config save_next_config(response, t=result_timestamp) # change config change_conf(response['recommendation']) @task def run_loops(max_iter=1): # dump database if it's not done before. dump = dump_database() for i in range(int(max_iter)): if dconf.RELOAD_INTERVAL > 0: if i % dconf.RELOAD_INTERVAL == 0: if i == 0 and dump is False: restart_database() restore_database() elif i > 0: restore_database() LOG.info('The %s-th Loop Starts / Total Loops %s', i + 1, max_iter) loop(i % dconf.RELOAD_INTERVAL if dconf.RELOAD_INTERVAL > 0 else i) LOG.info('The %s-th Loop Ends / Total Loops %s', i + 1, max_iter) @task def rename_batch(result_dir=None): result_dir = result_dir or dconf.RESULT_DIR results = glob.glob(os.path.join(result_dir, '*__summary.json')) results = sorted(results) for i, result in enumerate(results): prefix = os.path.basename(result) prefix_len = os.path.basename(result).find('_') + 2 prefix = prefix[:prefix_len] new_prefix = str(i) + '__' for base in ('summary', 'knobs', 'metrics_before', 'metrics_after'): fpath = os.path.join(result_dir, prefix + base + '.json') rename_path = os.path.join(result_dir, new_prefix + base + '.json') os.rename(fpath, rename_path) def _http_content_to_json(content): if isinstance(content, bytes): content = content.decode('utf-8') try: json_content = json.loads(content) decoded = True except (TypeError, json.decoder.JSONDecodeError): json_content = None decoded = False return json_content, decoded def _modify_website_object(obj_name, action, verbose=False, **kwargs): verbose = _parse_bool(verbose) if obj_name == 'project': valid_actions = ('create', 'edit') elif obj_name == 'session': valid_actions = ('create', 'edit') elif obj_name == 'user': valid_actions = ('create', 'delete') else: raise ValueError('Invalid object: {}. Valid objects: project, session'.format(obj_name)) if action not in valid_actions: raise ValueError('Invalid action: {}. Valid actions: {}'.format( action, ', '.join(valid_actions))) data = {} for k, v in kwargs.items(): if isinstance(v, (dict, list, tuple)): v = json.dumps(v) data[k] = v url_path = '/{}/{}/'.format(action, obj_name) response = requests.post(CONF['upload_url'] + url_path, data=data) content = response.content.decode('utf-8') if response.status_code != 200: raise Exception("Failed to {} new {}.\nStatus: {}\nMessage: {}\n".format( action, obj_name, response.status_code, content)) json_content, decoded = _http_content_to_json(content) if verbose: if decoded: LOG.info('\n%s_%s = %s', action.upper(), obj_name.upper(), json.dumps(json_content, indent=4)) else: LOG.warning("Content could not be decoded.\n\n%s\n", content) return response, json_content, decoded @task def create_website_user(**kwargs): return _modify_website_object('user', 'create', **kwargs) @task def delete_website_user(**kwargs): return _modify_website_object('user', 'delete', **kwargs) @task def create_website_project(**kwargs): return _modify_website_object('project', 'create', **kwargs) @task def edit_website_project(**kwargs): return _modify_website_object('project', 'edit', **kwargs) @task def create_website_session(**kwargs): return _modify_website_object('session', 'create', **kwargs) @task def edit_website_session(**kwargs): return _modify_website_object('session', 'edit', **kwargs) def wait_pipeline_data_ready(max_time_sec=800, interval_sec=10): max_time_sec = int(max_time_sec) interval_sec = int(interval_sec) elapsed = 0 while elapsed <= max_time_sec: response = requests.get(dconf.WEBSITE_URL + '/test/pipeline/') response = response.content LOG.info(response) if 'False' in str(response): time.sleep(interval_sec) elapsed += interval_sec else: return @task def integration_tests(): # Create test website response = requests.get(dconf.WEBSITE_URL + '/test/create/') LOG.info(response.content) # Upload training data LOG.info('Upload training data to no tuning session') upload_batch(result_dir='../../integrationTests/data/', upload_code='ottertuneTestNoTuning') # wait celery periodic task finishes wait_pipeline_data_ready() # Test DNN LOG.info('Test DNN (deep neural network)') upload_result(result_dir='../../integrationTests/data/', prefix='0__', upload_code='ottertuneTestTuningDNN') response = get_result(upload_code='ottertuneTestTuningDNN') assert response['status'] == 'good' # Test GPR LOG.info('Test GPR (gaussian process regression)') upload_result(result_dir='../../integrationTests/data/', prefix='0__', upload_code='ottertuneTestTuningGPR') response = get_result(upload_code='ottertuneTestTuningGPR') assert response['status'] == 'good'