Using LD_PRELOAD to run any version of SQLite with Python

I've been wanting to figure this out for ages. This thread on Hacker News plus this Stackoverflow post gave me some initial clues.

The challenge: someone reports a bug in Datasette with a specific version of SQLite. How can I run Datasette against that specific version?

Steps:

  1. Download the source code for that specific version
  2. Build a libsqlite3.so file using that source code
  3. Run Datasette prefixed with LD_PRELOAD=path/to/libsqlite3.so

Obtaining source code for old SQLite versions

This took a while to figure out. https://www.sqlite.org/chronology.html has a list of links to different releases. From there clicking on one of the dates takes you to a page like this: https://www.sqlite.org/src/timeline?c=cf538e2783&y=ci

From there clicking the "version-X" link takes you through to the timeline page like this: https://www.sqlite.org/src/timeline?r=version-3.8.11.1 - then click on e.g. "check-in: cf538e27" to get to a page like https://www.sqlite.org/src/info/cf538e2783e468bb which has a link to a Tarball and a Zip archive. Download one of those!

Hopefully shorter version: decide which version you want, then navigate to https://www.sqlite.org/src/timeline?t=version-3.8.11.1 and click the check-in link to get to the page that links to the Zip / Tarball.

Here's the URL to the 3.8.11.1 Tarball: https://www.sqlite.org/src/tarball/cf538e27/SQLite-cf538e27.tar.gz

Building the library

I first tried this on macOS but I couldn't get the LD_PRELOAD trick to work. I think this is because of "system integrity protection" (clue in this comment) - so I decided to do it in a Docker Ubuntu container instead.

To start a fresh Ubuntu container:

docker run -it -p 8001:8001 ubuntu

I'm using -p 8001:8001 here to allow me to access Datasette running on port 8001 later on.

Now we need python3 and the various build tools (SQLite also needs tcl for the build):

apt-get install -y python3 build-essential tcl wget

Annoyingly installing tcl requires us to interactively set the timezone data - I enter 2 for America and then 85 for Los Angeles.

Download and extract the SQLite code:

cd /tmp
wget https://www.sqlite.org/src/tarball/cf538e27/SQLite-cf538e27.tar.gz
tar -xzvf SQLite-cf538e27.tar.gz
cd SQLite-cf538e27

Now we can build the extension. The CPPFLAGS are optional but I found I needed them to get the full Datasette test suite to pass later on:

CPPFLAGS="-DSQLITE_ENABLE_FTS3 -DSQLITE_ENABLE_FTS3_PARENTHESIS -DSQLITE_ENABLE_RTREE=1" ./configure
make

There is no need to run make install here, we just need the compiled library.

This puts the libraries in a .libs directory:

root@11b79085483e:/tmp/SQLite-cf538e27# ls -l .libs/
total 22660
-rw-r--r-- 1 root root 9622164 Jun 17 12:52 libsqlite3.a
lrwxrwxrwx 1 root root      16 Jun 17 12:52 libsqlite3.la -> ../libsqlite3.la
-rw-r--r-- 1 root root     955 Jun 17 12:52 libsqlite3.lai
lrwxrwxrwx 1 root root      19 Jun 17 12:52 libsqlite3.so -> libsqlite3.so.0.8.6
lrwxrwxrwx 1 root root      19 Jun 17 12:52 libsqlite3.so.0 -> libsqlite3.so.0.8.6
-rwxr-xr-x 1 root root 4021816 Jun 17 12:52 libsqlite3.so.0.8.6
-rwxr-xr-x 1 root root  314160 Jun 17 12:52 sqlite3
-rw-r--r-- 1 root root 9232752 Jun 17 12:52 sqlite3.o

We can confirm that this worked using the following one-liner:

LD_PRELOAD=.libs/libsqlite3.so python3 -c \
    'import sqlite3; print(sqlite3.connect(":memory").execute("select sqlite_version()").fetchone())'

Which for me now outputs:

('3.8.11.1',)

Try that without the LD_PRELOAD to see the difference:

root@11b79085483e:/tmp/SQLite-cf538e27# python3 -c \
> 'import sqlite3; print(sqlite3.connect(":memory").execute("select sqlite_version()").fetchone())'
('3.31.1',)

Running Datasette

No need for a virtual environent since we are in a Docker container. We need to install pip:

apt-get install -y python3-pip

Then we can install Datasette:

python3 -mpip install datasette

Now run it like this:

datasette -h 0.0.0.0
INFO:     Started server process [10266]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:8001 (Press CTRL+C to quit)

We need -h 0.0.0.0 to allow access from outside. On our machine we can now visit it here:

http://0.0.0.0:8001/-/versions

And confirm that the default SQLite version is "version": "3.31.1".

Alternatively, run this to see the versions output without needing to run the server:

datasette --get /-/versions.json

Quit Datasette and start it again with the LD_PRELOAD:

LD_PRELOAD=.libs/libsqlite3.so datasette -h 0.0.0.0
INFO:     Started server process [10274]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:8001 (Press CTRL+C to quit)

And now http://0.0.0.0:8001/-/versions reports the following:

{
    "python": {
        "version": "3.8.2",
        "full": "3.8.2 (default, Apr 27 2020, 15:53:34) \n[GCC 9.3.0]"
    },
    "datasette": {
        "version": "0.44"
    },
    "asgi": "3.0",
    "uvicorn": "0.11.5",
    "sqlite": {
        "version": "3.8.11.1",
        "fts_versions": [],
        "extensions": {},
        "compile_options": [
            "HAVE_ISNAN",
            "SYSTEM_MALLOC",
            "TEMP_STORE=1",
            "THREADSAFE=1"
        ]
    }
}

Running the Datasette tests

To run Datasette's test suite I needed to install a few extra dependencies:

apt-get install -y python3-pip git python3.8-venv
cd /tmp
git clone https://github.com/simonw/datasette
cd datasette
python3 -m venv venv
source venv/bin/activate
pip install wheel # So bdist_wheel works in next step
pip install -e '.[test]'
LD_PRELOAD=/tmp/SQLite-cf538e27/.libs/libsqlite3.so python3 -c 

I used this process to track down a bug that was introduced between SQLite 3.36.0 and 3.37.0 in this issue.

Created 2020-06-17T13:12:13-07:00, updated 2022-03-05T11:55:51-08:00 · History · Edit