我正在尝试为法国公共药物数据库编写一个解析器/API(https://base-donnees-publique.medicaments.gouv.fr/)。它由八个 CSV 文件(实际上是 TSV,因为他们使用制表符)组成,每个文件从几 kB 到 4 MB,最大的有约 20000 行(每行代表一种药物及其名称、代码、价格等)。
由于这些文件可能会定期出现,我想直接解析它们,而不是创建一个更干净的数据库(因为我可能必须定期重新创建它)。
导入这些文件需要一点时间(大约一秒钟),所以我尝试加快一点速度并对几种方法进行了一些基准测试,我惊讶地发现最基本的方法似乎也是最快的。
这是我的测试代码(抱歉它很长)。每个文件都与一个专用类关联来解析其行。基本上,这些类是命名元组,具有自定义类方法来解析日期、数字等。
import pathlib
import enum
import datetime
from decimal import Decimal
from collections import namedtuple
import csv
def parse_date(date: str) -> datetime.datetime:
return datetime.datetime.strptime(date, "%d/%m/%Y").date()
def parse_date_bis(date: str) -> datetime.datetime:
return datetime.datetime.strptime(date, "%Y%m%d").date()
def parse_text(text):
if not text:
return ""
return text.replace("<br>", "\n").strip()
def parse_list(raw):
return raw.split(";")
def parse_price(price: str) -> Decimal:
if not price:
return None
# Handles cases like "4,417,08".
price = '.'.join(price.rsplit(",", 1)).replace(',', '')
return Decimal(price)
def parse_percentage(raw: str) -> int:
if not raw:
return None
return int(raw.replace("%", "").strip())
class StatutAdministratifPresentation(enum.Enum):
ACTIVE = "Présentation active"
ABROGEE = "Présentation abrogée"
class EtatCommercialisation(enum.Enum):
DC = "Déclaration de commercialisation"
S = "Déclaration de suspension de commercialisation"
DAC = "Déclaration d'arrêt de commercialisation"
AC = "Arrêt de commercialisation (le médicament n'a plus d'autorisation)"
class MotifAvisSMR(enum.Enum):
INSCRIPTION = "Inscription (CT)"
RENOUVELLEMENT = "Renouvellement d'inscription (CT)"
EXT = "Extension d'indication"
EXTNS = "Extension d'indication non sollicitée"
REEV_SMR = "Réévaluation SMR"
REEV_ASMR = "Réévaluation ASMR"
REEV_SMR_ASMR = "Réévaluation SMR et ASMR"
REEV_ETUDE = "Réévaluation suite à résultats étude post-inscript"
REEV_SAISINE = "Réévaluation suite saisine Ministères (CT)"
NOUV_EXAM = "Nouvel examen suite au dépôt de nouvelles données"
MODIF_COND = "Modification des conditions d'inscription (CT)"
AUTRE = "Autre demande"
class ImportanceSMR(enum.Enum):
IMPORTANT = "Important"
MODERE = "Modéré"
FAIBLE = "Faible"
INSUFFISANT = "Insuffisant"
COMMENTAIRES = "Commentaires"
NP = "Non précisé"
class ImportanceASMR(enum.Enum):
COM = "Commentaires sans chiffrage de l'ASMR"
I = "I"
II = "II"
III = "III"
IV = "IV"
V = "V"
NP = "Non précisée"
SO = "Sans objet"
class Specialite(namedtuple("Specialite", ("cis", "denomation", "forme", "voies_administration", "statut_amm", "type_amm", "commercialisation", "date_amm", "statut_bdm", "numero_autorisation_europeenne", "titulaire", "surveillance_renforcee"))):
@classmethod
def from_line(cls, line):
line[2] = line[2].replace(" ", " ").strip()
line[3] = parse_list(line[3])
line[7] = parse_date(line[7])
line[10] = line[10].strip() # There are often leading spaces here (like ' OPELLA HEALTHCARE FRANCE').
return cls(*line)
class Presentation(namedtuple("Specialite", ("cis", "cip7", "libelle", "statut", "commercialisation", "date_commercialisation", "cip13", "agrement_collectivites", "taux_remboursement", "prix", "prix_hors_honoraires", "montant_honoraires", "indications_remboursement"))):
@classmethod
def from_line(cls, line):
if line[3] == "Présentation active":
line[3] = StatutAdministratifPresentation.ACTIVE
else:
line[3] = StatutAdministratifPresentation.ABROGEE
line[4] = {
"Déclaration de commercialisation": EtatCommercialisation.DC,
"Déclaration de suspension de commercialisation": EtatCommercialisation.S,
"Déclaration d'arrêt de commercialisation": EtatCommercialisation.DAC,
"Arrêt de commercialisation (le médicament n'a plus d'autorisation)": EtatCommercialisation.AC
}.get(line[4])
line[5] = parse_date(line[5])
line[7] = True if line[7] == "oui" else False
line[8] = parse_percentage(line[8])
line[9] = parse_price(line[9])
line[10] = parse_price(line[10])
line[11] = parse_price(line[11])
line[12] = parse_text(line[12])
return cls(*line)
class Composition(namedtuple("Composition", ("cis", "element", "code", "substance", "dosage", "ref_dosage", "nature_composant", "cle"))):
@classmethod
def from_line(cls, line):
line.pop(-1)
return cls(*line)
class AvisSMR(namedtuple("AvisSMR", ("cis", "dossier_has", "motif", "date", "valeur", "libelle"))):
@classmethod
def from_line(cls, line):
line[2] = MotifAvisSMR(line[2])
line[3] = parse_date_bis(line[3])
line[4] = ImportanceSMR(line[4])
line[5] = parse_text(line[5])
return cls(*line)
class AvisASMR(namedtuple("AvisASMR", ("cis", "dossier_has", "motif", "date", "valeur", "libelle"))):
@classmethod
def from_line(cls, line):
line[2] = MotifAvisSMR(line[2])
line[3] = parse_date_bis(line[3])
line[4] = ImportanceASMR(line[4])
line[5] = parse_text(line[5])
return cls(*line)
class AvisCT(namedtuple("AvisCT", ("dossier_has", "lien"))):
@classmethod
def from_line(cls, line):
return cls(*line)
FILE_MATCHES = {
"CIS_bdpm.txt": Specialite,
"CIS_CIP_bdpm.txt": Presentation,
"CIS_COMPO_bdpm.txt": Composition,
"CIS_HAS_ASMR_bdpm.txt": AvisASMR,
"CIS_HAS_SMR_bdpm.txt": AvisSMR,
"HAS_LiensPageCT_bdpm.txt": AvisCT
}
def sequential_import_file_data(filename, cls):
result = {cls: []}
with (pathlib.Path("data") / filename).open("r", encoding="latin1") as f:
rows = csv.reader(f, delimiter="\t")
for line in rows:
data = cls.from_line(line)
result[cls].append(data)
return result
def import_data_sequential():
results = []
for filename, cls in FILE_MATCHES.items():
results.append(sequential_import_file_data(filename, cls))
from multiprocessing.pool import ThreadPool
def import_data_mp_tp(n=2):
pool = ThreadPool(n)
results = []
for filename, cls in FILE_MATCHES.items():
results.append(pool.apply_async(
sequential_import_file_data,
(filename, cls)
))
results = [r.get() for r in results]
from multiprocessing.pool import Pool
def import_data_mp_p(n=2):
pool = Pool(n)
results = []
for filename, cls in FILE_MATCHES.items():
results.append(pool.apply_async(
sequential_import_file_data,
(filename, cls)
))
results = [r.get() for r in results]
import asyncio
import aiofiles
from aiocsv import AsyncReader
async def async_import_file_data(filename, cls):
results = {cls: []}
async with aiofiles.open(
(pathlib.Path("data") / filename),
mode="r",
encoding="latin1"
) as afp:
async for line in AsyncReader(afp, delimiter="\t"):
data = cls.from_line(line)
results[cls].append(data)
return results
def import_data_async():
results = []
for filename, cls in FILE_MATCHES.items():
results.append(asyncio.run(async_import_file_data(filename, cls)))
def main():
import timeit
print(
"Sequential:",
timeit.timeit(lambda: import_data_sequential(), number=10)
)
print(
"Multi ThreadPool:",
timeit.timeit(lambda: import_data_mp_tp(), number=10)
)
print(
"Multi Pool:",
timeit.timeit(lambda: import_data_mp_p(), number=10)
)
print(
"Async:",
timeit.timeit(lambda: import_data_async(), number=10)
)
if __name__ == "__main__":
main()
所以当我运行它时,我得到以下结果。
Sequential: 9.821639589001279
Multi ThreadPool: 10.137484730999859
Multi Pool: 12.531487682997977
Async: 30.953154197999538
迭代所有文件及其所有行的最基本解决方案似乎也是最快的。
那么,我是否做错了什么会减慢导入速度?或者说有这样的时差是正常/预期的吗?
编辑2023-08-15:我意识到,由于我需要解析的所有文件都是TSV(并且它们的值中不包含制表符),我仍然可以通过使用简单的
line.strip('\n').split('\t')
来加快解析速度而不是 CSV 模块,运行时间又节省了 40%。 :) 当我有这个数据库的完整 API 时,我可能会发布一个要点。
像往常一样:在代码上运行分析器以查看其时间花在哪里。 (这是 PyCharm 的,它包装了 stdlib
cProfile
。)
顺序:7.865187874995172
嗯,好吧。
strptime
,我知道会被datetime.datetime.strptime
打电话。另外,奇怪的是,getlocale
...为什么我们需要那里的语言环境?单击调用图显示 strptime
实际上查找当前区域设置,并且有一堆锁等等 - 如果我们用我们自己的实现替换这些 parse_date
会怎么样?
def parse_date(date: str) -> datetime.date:
d, m, y = (int(x) for x in date.split("/", 2))
return datetime.date(2000 + y, m, d)
def parse_date_bis(date: str) -> datetime.datetime:
y = int(date[:4])
m = int(date[4:6])
d = int(date[6:8])
return datetime.datetime(y, m, d)
顺序:3.8978060420195106
好的,我们做饭了! 52% 的改进就在那里!
(它没有出现在此处的屏幕截图中,因为我是一只愚蠢的鹅裁剪它,但是
re
在引擎盖下使用的 strptime
东西也立即掉落了。)
现在让我们假设在那些热门
@lru_cache(maxsize=None)
函数上会有很多相同的日期和 slap parse_date_*
(RAM 灵活,无界缓存),运行代码并打印出缓存信息:
Sequential: 3.2240814580000006
CacheInfo(hits=358989, misses=6991, maxsize=None, currsize=6991)
CacheInfo(hits=221607, misses=513, maxsize=None, currsize=513)
我觉得还不错,我们最后一个号码还有 15% 的折扣。
不过,parse_price
显然也可以使用缓存:
Sequential: 2.928746833000332
CacheInfo(hits=358989, misses=6991, maxsize=None, currsize=6991)
CacheInfo(hits=221607, misses=513, maxsize=None, currsize=513)
CacheInfo(hits=622064, misses=4096, maxsize=None, currsize=4096)
嘿,谁知道,数据中只有 4096 个单独的价格字符串。
如果你有足够的内存,剩下的解析函数也可以使用缓存,但是通过一点点分析和解析苦劳,它现在的速度提高了 2.7 倍 [当运行所有内容 10 次时,这意味着这些缓存将会很热 – 单个run 的加速并不那么显着],不需要并行处理。魔法!
为了让比赛场地更加均匀,这里有一个
hyperfine
基准测试,其中 Python 解释器对于每次导入都从头开始(每个解释器仅运行一次导入):
$ hyperfine 'python3 so76781391-orig.py' 'python3 so76781391-opt.py' --warmup 5 --min-benchmarking-time 10
Benchmark 1: python3 so76781391-orig.py
Time (mean ± σ): 363.0 ms ± 2.7 ms [User: 340.8 ms, System: 20.7 ms]
Range (min … max): 358.9 ms … 367.9 ms 27 runs
Benchmark 2: python3 so76781391-opt.py
Time (mean ± σ): 234.1 ms ± 2.5 ms [User: 215.6 ms, System: 17.0 ms]
Range (min … max): 228.2 ms … 238.5 ms 42 runs
Summary
'python3 so76781391-opt.py' ran
1.55 ± 0.02 times faster than 'python3 so76781391-orig.py'
所以,通过快速查看分析器,具体速度提升了 55%(以及一些额外的优化,例如不在
from_line
函数中创建映射字典等)。