From 915027f60745a9765af46e75a04e62814c865e2f Mon Sep 17 00:00:00 2001
From: Wojciech Sromek <157375010+wojciechsromek@users.noreply.github.com>
Date: Tue, 14 Jan 2025 18:04:17 +0100
Subject: [PATCH] doc: Provide ADR for load performance optimisation (#936)

---
 adr/002-load-performance-optimisation.md | 31 ++++++++++++++++++++++++
 1 file changed, 31 insertions(+)
 create mode 100644 adr/002-load-performance-optimisation.md

diff --git a/adr/002-load-performance-optimisation.md b/adr/002-load-performance-optimisation.md
new file mode 100644
index 000000000..f25e83354
--- /dev/null
+++ b/adr/002-load-performance-optimisation.md
@@ -0,0 +1,31 @@
+# ADR-002: Load Performance Optimisation
+
+### Links
+[Notion Page on the boot optimisation](https://www.notion.so/worldcoin/SMPCv2-Boot-time-optimisation-1458614bdf8c80c39170dc0040eb9acf)
+
+## Context
+- on every boot-up the container needs to load a couple of 100s of GBs (and the number is constantly growing, hopefully into TBs)
+- the load is performed directly from the AuroraDB 
+- we need to be able to boot the container below 10 minutes
+
+## Decision
+We have decided to prepare an `iris-mpc-db-exporter` service that is going to be periodically storing a dump of the database in an S3 bucket. The loading of the data from the bucket, based on our research are going to be significantly faster than from the AuroraDB.   
+
+Aiming to boost the performance further we have decided to move the conversion of `u8` into `i8` to the exporting service. 
+
+**What was also implemented there was a modification of the order of the bytes for the sake of faster loading into the memory of `iris-mpc` pods.**
+
+## Rationale
+- The loading of the data from the S3 bucket is significantly faster than from the AuroraDB.
+- The conversion of `u8` into `i8` is a simple operation that can be done in the exporting service.
+- The modification of the order of the bytes is a simple operation that can be done in the exporting service.
+
+## Consequences
+Because of the actions required to make the loading of the data more efficient (altered byte order, byte values shifted) right now we have two storages:
+- AuroraDB (source of truth)
+- S3 bucket (used for improving the data load time)
+
+They both store the same data **but in different formats**, please refer to implementation (references are accurate in revision `ec14a0c022b73d2d291578502f6bf02cbc0a99d0`):
+- `iris-mpc-store/src/lib.rs:41` (universal type for both data formats)
+- `iris-mpc-gpu/src/server/actor.rs:464` (method for loading data from AuroraDB into memory)
+- `iris-mpc-gpu/src/server/actor.rs:503` (method for loading data from S3 bucket into memory)