Crossing 2,500 Discord guilds mandates WebSocket Gateway sharding, while Lavalink v4 must pace 20 ms UDP Opus audio packets without stop-the-world JVM garbage collection hiccups. On BotFoundry, we pre-touch fixed -Xms/-Xmx heaps under Java 21 Generational ZGC and cap Discord.js guild caches.
Public rates retrieved 2026-11-12 from official vendor pricing pages cited below; promotional & renewal prices may change — always verify exact specs, term length, and renewal rates on the vendor’s official checkout page.
| Concurrency Scale | vCPU | DDR5 RAM | OS / Runtime | /dev/shm & Watchdog Guard | Public Tier Match (2026-11-12) |
|---|---|---|---|---|---|
| 1–2 Shards + 15 Audio Streams | 2 vCPU KVM | 4 GB DDR5 | Ubuntu 24.04 + Java 21 | ZGC -Xmx2G + RestartSec=3 |
Hostinger KVM 1 / Vultr HF 4GB |
| 4–8 Shards + 60 Audio Streams | 4 vCPU 3.5GHz+ HF | 8 GB DDR5 | Ubuntu 24.04 ZGC | ZGC -Xmx4G + AlwaysPreTouch |
Vultr HF 8GB / Hostinger KVM 2 |
| 16–32 Shards + 200 Streams | 6–8 Dedicated VDS | 16 GB DDR5 | Ubuntu 24.04 + Redis | ZGC -Xmx8G + Swappiness=1 |
UltaHost VDS 16GB / Hostinger KVM 4 |
| 64+ Shards + Multi-Region Audio | 12 Core Ryzen / VDS | 32–64 GB DDR5 | Clustered Lavalink Nodes | IPv6 /64 Planner + VDS |
InterServer Ryzen / UltaHost VDS 32GB |
[Service]
User=lavalink
WorkingDirectory=/opt/lavalink
ExecStart=/usr/bin/java \
-Xms4G -Xmx4G \
-XX:+UseZGC -XX:+ZGenerational \
-XX:+AlwaysPreTouch \
-Djdk.nio.maxCachedBufferSize=262144 \
-jar Lavalink.jar
MemoryMax=5G
Restart=always
RestartSec=3Legacy G1GC stop-the-world pauses can spike past 25–40 ms when dozens of voice channels skip tracks simultaneously, interrupting 20 ms Opus frame pacing. Enabling -XX:+UseZGC -XX:+ZGenerational with matching -Xms4G -Xmx4G and -XX:+AlwaysPreTouch keeps GC pauses sub-millisecond.
Because Discord’s primary Gateway API operates in US-East, placing your Shard Manager in New Jersey/New York ensures fast 3-second Slash Command acknowledgments, while optional regional Lavalink satellites in Singapore or Frankfurt serve local voice channels.
Storing unlimited member presences and message histories across 10,000 guilds leaks 10+ GB of Node.js heap; enforce Options.cacheWithLimits() and omit GUILD_PRESENCES unless strictly required.
With direct Opus frame passthrough (no heavy real-time DSP equalization), 1 high-clock vCPU handles ~25–35 active streams; with real-time karaoke/timescale DSP filters active, budget 8–12 streams per vCPU.
Attach a routed IPv6 /64 subnet (available on Vultr and InterServer) with Lavalink’s RotatingNanoIpRoutePlanner so outbound requests cycle cleanly across addresses.
JVM processes consume native off-heap memory for thread stacks, Netty NIO direct buffers, and native Opus C libraries on top of the 4 GB Java heap — leaving a 1 GB off-heap margin prevents premature cgroup OOM kills.
BotFoundry 4-layer RAM equation, /portable chart trimming, and Windows AutoLogon watchdog rules for 1 to 50+ MetaTrader 4/5 Expert Advisor terminals.
BotFoundry sizing formulas for Ubuntu 24.04 Freqtrade Docker bots, cgroup MemoryMax bounds, chronyd NTP sync, and isolated FreqAI LightGBM retraining.
BotFoundry memory math for headless Chromium BrowserContexts, Docker --shm-size=2g shared memory pools, tini zombie reaping, and flat-bandwidth KVM nodes.
BotFoundry 4-layer RAM guide for scaling self-hosted n8n Queue Mode with dedicated webhook receivers, Redis AOF queues, and PostgreSQL 16 pruning.