Paxos

Disclaimer

In my implementation I use a non-blocking UDP socket. During my testing in a VM, I encountered a buffer overflow, in order for the code to work on any host and number of messages. I introduced a timeout of 0.1 seconds (tested up to 10'000 messages including message loss) between each of the clients messages to the proposer. This can be set in the paxos.conf file as follows:

clients  239.0.0.1 5000
proposers 239.0.0.1 6000
acceptors 239.0.0.1 7000
learners   239.0.0.1 8000
timeout 0.1

Please keep this in mind when testing my implementation and if necessary increase the timeout until paxos is killed.

Dependencies

You need bash and basic unix tools (grep, sed, etc). If you want to run the test with msg losses, you need iptables with sudo access.

Depending on the network interface used, ip multicast might not be enabled. You can check using "ifconfig" and checking that the "MULTICAST" flag is set. You might have to enable it using:

ifconfig IFACE multicast

where IFACE is the name of the interface. Using a connected cable/wifi interface probably will not have this problem (e.g. "eth0", "wlan0").

Python

Make sure you have python12 or higher installed and use one of the following two options to install the dependencies:

Option 1: I only added a single package to the project, which is called loguru and can be installed with:

pip install loguru

Option 2: I used the Poetry package manager for this python project and you can replicate my setup by using the following commands:

# Install all the dependencies
poetry install

# Enter virtual env
poetry shell

Logging output

By using loguru you can control the level of output as follows

# Without debugging messages
export LOGURU_LEVEL="INFO"

# Including debug messages
export LOGURU_LEVEL="DEBUG"

Run Program

In term of structure it is identical to the one provided by the project. You can run the scrits inside the paxos folder as follows

./run.sh fake-paxos 100
./check_all.sh

Timeouts

I introduce three different timeouts in my implementation:

As described in the Disclaimer the client sends in the define interval each message to the proposer.
The learner tries to catch up every second, if there is an instance number missing, which is smaller than the highest intance number received
The Proposer tries to repropose a message after one or two seconds (chosen randomly), if no consensus was reached for the clients value.

How to run the tests

Place the root of your paxos implementation inside this directory. For example, if you have the folder ~/paxos-tests/MyPaxos, your *.sh scripts should be directly inside it (e.g. ~/paxos-tests/MyPaxos/acceptor.sh) and should work when called from inside the directory itself.
Run one of the run*.sh from inside THIS folder. When the run finishes, run check_all.sh to check the output of the run. For example:

cd ~/paxos-tests/
./run.sh MyPaxos 100  # each client will submit 100 values
# wait for the run to finish
./check_all.sh # check the run

After a run ends, run check_all.sh to see if everything went OK. "Test 3" might FAIL in some cases, but with few proposed values and no message loss it should also be OK.

clennys/Paxos