Reproduction steps

1. Perform database setup and get dependencies:

./script/setup.sh

2. Start message consumer process:

./start-service.sh

3. Start one of message producers:

./script/produce-messages_1_long_transactions.sh
./script/produce-messages_2_batches_in_contention.sh
./script/produce-messages_3_message_in_contention.sh

4. Verify if number of Deposit messages equals number of Deposited. If not, you've experienced race condition in get_category_messages

psql message_store < check_messages.sql

Reproduction on video

The example illustrating long transactions (highest chance of race condition):

• https://v.usetapes.com/fFFLdpUS2t

The example illustraing uneven but short transactions (batches of 1-2 messages):

• https://v.usetapes.com/mipFBvNo8F
• https://v.usetapes.com/bAzuKpJQYr

The example illustrating most typical usage under high load:

• https://v.usetapes.com/gYxbsabA3D

The problem exposed by the experiments:

In such case we can get IDs of events 0 and 2, but skip reading uncommitted event nr 1.