SQLite 版本引发的 Python 程序调用问题

问题

在跑 OpenStack functional 功能测试的时候有两个用例过不去。

• nova.tests.functional.db.test_resource_provider.ResourceClassTestCase.test_create_duplicate_id_retry
• nova.tests.functional.db.test_resource_provider.ResourceClassTestCase.test_create_duplicate_id_retry_failing

调试定位到问题代码：

# /opt/stack/queens/nova/nova/objects/resource_provider.py

    def create(self):
        if 'id' in self:
            raise exception.ObjectActionError(action='create',
                                              reason='already created')
        if 'name' not in self:
            raise exception.ObjectActionError(action='create',
                                              reason='name is required')
        if self.name in fields.ResourceClass.STANDARD:
            raise exception.ResourceClassExists(resource_class=self.name)

        if not self.name.startswith(fields.ResourceClass.CUSTOM_NAMESPACE):
            raise exception.ObjectActionError(
                action='create',
                reason='name must start with ' +
                        fields.ResourceClass.CUSTOM_NAMESPACE)

        updates = self.obj_get_changes()
        # There is the possibility of a race when adding resource classes, as
        # the ID is generated locally. This loop catches that exception, and
        # retries until either it succeeds, or a different exception is
        # encountered.
        retries = self.RESOURCE_CREATE_RETRY_COUNT
        while retries:
            retries -= 1
            try:
                rc = self._create_in_db(self._context, updates)
                self._from_db_object(self._context, self, rc)
                break
            except db_exc.DBDuplicateEntry as e:
                # NOTE: e.columns 为空，所以直接出发后续的异常
                if 'id' in e.columns:
                    # Race condition for ID creation; try again
                    continue
                # The duplication is on the other unique column, 'name'. So do
                # not retry; raise the exception immediately.
                raise exception.ResourceClassExists(resource_class=self.name)
        else:
            # We have no idea how common it will be in practice for the retry
            # limit to be exceeded. We set it high in the hope that we never
            # hit this point, but added this log message so we know that this
            # specific situation occurred.
            LOG.warning("Exceeded retry limit on ID generation while "
                        "creating ResourceClass %(name)s",
                        {'name': self.name})
            msg = _("creating resource class %s") % self.name
            raise exception.MaxDBRetriesExceeded(action=msg)

继续看 db_exc.DBDuplicateEntry 的实现：

# /usr/lib/python2.7/site-packages/oslo_db/exception.py
class DBDuplicateEntry(DBError):
    """Duplicate entry at unique column error.

    Raised when made an attempt to write to a unique column the same entry as
    existing one. :attr: `columns` available on an instance of the exception
    and could be used at error handling::

       try:
           instance_type_ref.save()
       except DBDuplicateEntry as e:
           if 'colname' in e.columns:
               # Handle error.

    :kwarg columns: a list of unique columns have been attempted to write a
        duplicate entry.
    :type columns: list
    :kwarg value: a value which has been attempted to write. The value will
        be None, if we can't extract it for a particular database backend. Only
        MySQL and PostgreSQL 9.x are supported right now.
    """
    def __init__(self, columns=None, inner_exception=None, value=None):
        # 正常情况下，触发 DBDuplicateEntry 会将冲突的 columns 返回，让开发者得以方便的作出进一步判断
        self.columns = columns or []
        self.value = value
        super(DBDuplicateEntry, self).__init__(inner_exception)

定位到生成冲突 columns 的地方：

# /opt/stack/queens/nova/.tox/functional/lib/python2.7/site-packages/oslo_db/sqlalchemy/exc_filters.py

@filters("sqlite", sqla_exc.IntegrityError,
         (r"^.*columns?(?P<columns>[^)]+)(is|are)s+nots+unique$",
          r"^.*UNIQUEs+constraints+failed:s+(?P<columns>.+)$",
          r"^.*PRIMARYs+KEYs+musts+bes+unique.*$"))
def _sqlite_dupe_key_error(integrity_error, match, engine_name, is_disconnect):
    """Filter for SQLite duplicate key error.

    note(boris-42): In current versions of DB backends unique constraint
    violation messages follow the structure:

    sqlite:
    1 column - (IntegrityError) column c1 is not unique
    N columns - (IntegrityError) column c1, c2, ..., N are not unique

    sqlite since 3.7.16:
    1 column - (IntegrityError) UNIQUE constraint failed: tbl.k1
    N columns - (IntegrityError) UNIQUE constraint failed: tbl.k1, tbl.k2

    sqlite since 3.8.2:
    (IntegrityError) PRIMARY KEY must be unique

    """
    columns = []
    # NOTE(ochuprykov): We can get here by last filter in which there are no
    #                   groups. Trying to access the substring that matched by
    #                   the group will lead to IndexError. In this case just
    #                   pass empty list to exception.DBDuplicateEntry
    try:
        columns = match.group('columns')
        columns = [c.split('.')[-1] for c in columns.strip().split(", ")]
    except IndexError:
        pass

    raise exception.DBDuplicateEntry(columns, integrity_error)

没有生产冲突 columns 的原因是：「底层 db engine 返回的 string match 不符合上述的匹配规范」。e.g.

2013-05-20 错误：('(sqlite3.IntegrityError) PRIMARY KEY must be unique',)
2019-04-16 正确：('(sqlite3.IntegrityError) UNIQUE constraint failed: resource_classes.id',)

这是一个 SQLite3 版本不匹配导致的问题，但在 Nova 项目中却没有明确的指定 SQLite3 的版本，所以只能手动的修复这一问题。

解决

手动编译升级 SQLite3 的版本：

wget https://www.sqlite.org/2019/sqlite-autoconf-3280000.tar.gz
tar -xvf sqlite-autoconf-3280000.tar.gz
cd sqlite-autoconf-3280000
mkdir /opt/sqlite3
./configure --prefix=/opt/sqlite3
make && make install

升级完 SQLite3 依旧没有直接解决上述问题，这里主要涉及到一个 Python 如何调用 C so 库的问题，这也是解决这个问题的精髓所在。

• 首先我们找到 SQLite3 Python 客户端（API）的位置

$ python -c "import sqlite3; print(sqlite3.__file__)"
/usr/lib64/python2.7/sqlite3/__init__.pyc

• 查看 SQLite3 API 实现并找到 so 库导入语句

# /usr/lib64/python2.7/sqlite3/dbapi2.py

from _sqlite3 import *

• 查找 _sqlite3 so 库的位置

$ python -c 'import _sqlite3; print(_sqlite3)'
<module '_sqlite3' from '/opt/stack/queens/nova/.tox/functional/lib64/python2.7/lib-dynload/_sqlite3.so'>

• 查看 _sqlite3 so 库内含的动态函数库

$ ldd /opt/stack/queens/nova/.tox/functional/lib64/python2.7/lib-dynload/_sqlite3.so
	linux-vdso.so.1 =>  (0x00007ffc4defb000)
	libsqlite3.so.0 => /lib64/libsqlite3.so.0 (0x00007f708ba42000)
	libpython2.7.so.1.0 => /lib64/libpython2.7.so.1.0 (0x00007f708b676000)
	libpthread.so.0 => /lib64/libpthread.so.0 (0x00007f708b45a000)
	libc.so.6 => /lib64/libc.so.6 (0x00007f708b08d000)
	libz.so.1 => /lib64/libz.so.1 (0x00007f708ae77000)
	libm.so.6 => /lib64/libm.so.6 (0x00007f708ab75000)
	libdl.so.2 => /lib64/libdl.so.2 (0x00007f708a971000)
	libutil.so.1 => /lib64/libutil.so.1 (0x00007f708a76e000)
	/lib64/ld-linux-x86-64.so.2 (0x00007f708bf62000)

• 凭直觉，我们首先关注 libsqlite3.so.0 函数库

$ ls -alh /lib64/libsqlite3.so.0
lrwxrwxrwx. 1 root root 19 May 14 05:13 /lib64/libsqlite3.so.0 -> libsqlite3.so.0.8.6

$ ls -alh /lib64/libsqlite3.so.0.8.6
-rwxr-xr-x. 1 root root 5.1M Jun  4 05:51 /lib64/libsqlite3.so.0.8.6

至此，我们可以想到之所以升级了 SQLite3 的版本但依旧没有解决问题的原因是「Python 程序中调用的动态函数库依旧没有被更新」。所以我们只需要使用新安装的 so 文件替换掉就的就可以解决了。

mv /usr/lib64/libsqlite3.so.0.8.6 /usr/lib64/libsqlite3.so.0.8.6.bk
cp /opt/sqlite3/lib/libsqlite3.so.0.8.6 /usr/lib64/libsqlite3.so.0.8.6

最后

最后贴上 SQLite3 的修改 commit：

This issue is involved this commit, and introduced by version-3.8.2

...
commit eb743f01b125bebd8736ceb2873b69f27721b0ae
Author: D. Richard Hipp <drh@hwaci.com>
Date:   Tue Nov 5 13:33:55 2013 +0000

    Standardize the error messages generated by constraint failures to a format
    of "$TYPE constraint failed: $DETAIL".  This involves many changes to the
    expected output of test cases.
...

解决这一问题主要的收获是 Python 程序和 C 程序之间的调用关键，如果两者之间并非是通过 TCP 协议来通信，而是通过 so 库文件来调用的话，那么我们需要注意 C 程序在 Linux 操作系统上的文件安装方式。并非单纯的升级了 C 程序就会立马在 Python 程序上生效，还要注意两者之间的桥梁（调用库文件）是否也一同升级了。