From 04fa11b996cb726e8942641ef839cbd623b4bd14 Mon Sep 17 00:00:00 2001 From: Tim Burns <92951498+OpenSource-For-Freedom@users.noreply.github.com> Date: Sun, 11 Oct 2026 17:45:23 -0400 Subject: [PATCH] Size the server for the 2 vCPU / 4 GB box Downsizing to cut the bill from $48 to $24. The pinned limits here were sized for 4 vCPU / 8 GB and three of them break on the smaller plan, one of them fatally: cpus was 3.5. Docker REFUSES to create a container whose cpus exceeds the host core count ("Range of CPUs is from 0.01 to 2.00, as there are only 2 CPUs available"), so on a 2 vCPU box this is not a performance question. compose up fails, the health check fails, and deploy.sh rolls back to a revision carrying the same bad value, leaving the server down until someone edits this line. Now 1.75: the same 87.5% share of the box that 3.5-of-4 was, keeping the headroom for sshd, tailscaled, auditd and the deploy timer that the old comment here argued for. mem_limit 6.5g and a 5G heap exceed the whole 4 GB host, so the kernel OOM-killer would take the server down under load. Now a 3.2g cap with a 2G heap, keeping about 1.2g for the JVM's non-heap needs and 800Mi for the host. These numbers are deliberately also safe on an 8 GB host, just wasteful there: a limit that fits the small plan cannot break the big one, while a limit sized for the big plan takes the small one down outright. MAX_PLAYERS goes 20 to 5. 20 was typed, never measured and never tested. A slot only costs memory while someone is in it, but the worst case is what has to fit, and on 4 GB it does not. Five is the real population. VIEW_DISTANCE 12 to 8 and SIMULATION_DISTANCE 8 to 6. Loaded chunks scale with the square of view distance, so this is about 56% fewer chunks held per player and the biggest lever on heap and tick time that costs no content. Simulation 6 still exceeds the 4-chunk radius that keeps a player's own farms and Create contraptions ticking. Be honest about the heap: 2G is under-provisioned for roughly 30 mods including Create, CC:Tweaked, BlueMap, Terralith, When Dungeons Arise and Ad Astra's five extra dimensions. It ran on 5G. Expect GC pressure. spark is in the mod list now, so this is measurable rather than guesswork, and the first cut if it will not hold is BlueMap: its port is bound to 127.0.0.1 and reachable only through an SSH tunnel, so no player can see what it spends memory rendering. Cutting mods that place blocks (Ad Astra, Create) is not an option, because removing them deletes what players built with them. --- docker-compose.yml | 86 ++++++++++++++++++++++++++++++---------------- 1 file changed, 57 insertions(+), 29 deletions(-) diff --git a/docker-compose.yml b/docker-compose.yml index e268cbb..9f7988b 100644 --- a/docker-compose.yml +++ b/docker-compose.yml @@ -21,17 +21,26 @@ services: # unlike a deploy.resources block. pids_limit caps processes (fork-bomb # guard); it can't live alongside a deploy.resources.limits block without # Compose flagging a phantom pids conflict. - mem_limit: 6.5g - mem_reservation: 4g - # Raised from 2.0 when the droplet went from 2 to 4 vCPUs (2026-08-07). - # This cap, not the droplet size, is what actually limits the server: the - # container sat pegged at ~197% of 2 cores during worldgen, and leaving - # this at 2.0 would have made the resize buy exactly nothing. - # 3.5, not 4.0: the host still needs cycles for sshd/tailscaled, auditd, - # the deploy timer and Docker itself. Handing the whole box to the JVM - # makes the admin plane unresponsive precisely when the server is busy, - # which is when you most need to get in. - cpus: 3.5 + # + # Sized for the SMALL box: 2 vCPU / 4 GB (s-2vcpu-4gb), the 2026-10-11 + # downsize from s-4vcpu-8gb to halve the bill from $48 to $24. These + # numbers are deliberately also safe on an 8 GB host, just wasteful there: + # a limit that fits the small plan cannot break the big one, while a limit + # sized for the big plan takes the small one down outright. + mem_limit: 3.2g + mem_reservation: 2g + # Docker REFUSES to create a container whose cpus exceeds the host core + # count: "Range of CPUs is from 0.01 to 2.00, as there are only 2 CPUs + # available". So a stale 3.5 on a 2 vCPU box is not a performance + # question, it is a hard failure - compose up fails, the health check + # fails, and the rollback restores the same bad value, leaving the server + # down until someone edits this line. + # 1.75, not 2.0: same 87.5% share of the box that 3.5-of-4 was, keeping + # the headroom the old comment here argued for. The host still needs + # cycles for sshd/tailscaled, auditd, the deploy timer and Docker itself, + # and handing the whole box to the JVM makes the admin plane unresponsive + # precisely when the server is busy, which is when you need to get in. + cpus: 1.75 pids_limit: 512 ulimits: nofile: @@ -121,8 +130,9 @@ services: # local DROP rule and upstream filtering look identical to a port scan. # So: do not remove a mod from this list because a client shows a red X. # Verify reachability first. Chunky can be re-added whenever wanted. - # The CPU change from that same PR is the part that mattered and it stays: - # cpus was raised 2.0 -> 3.5 after the droplet went to 4 vCPU. + # The CPU change from that same PR followed the hardware and is now + # undone: cpus went 2.0 -> 3.5 for 4 vCPU, then to 1.75 when the box was + # downsized to 2 vCPU on 2026-10-11. # # When Dungeons Arise (2.1.58) adds very large hand-built structures: # pirate ships, mountain castles, sky keeps, undead pyramids. Verified @@ -166,19 +176,25 @@ services: REMOVE_OLD_MODS: "TRUE" # Raised for Ad Astra, which adds five dimensions (Moon, Mars, Venus, # Mercury, Glacio) on top of Terralith and When Dungeons Arise. - # 5G heap inside a 6.5g container cap, NOT 6G. A JVM needs meaningful - # memory outside its heap: metaspace, code cache, GC structures, thread - # stacks and direct byte buffers. Setting the heap equal to the container - # limit is the classic way to get the kernel OOM-killer to kill the server - # under load, which looks exactly like a random crash. 1.5g of headroom is - # the margin that prevents it. - # The host has 7.8Gi total, so a 6.5g container leaves ~1.3Gi for the - # kernel, Docker, sshd, tailscaled, auditd and the deploy timer. Going - # higher starts trading the ability to log in for server memory, and this - # box has been locked out enough times already. - # A genuine 6G heap wants a 16 GB droplet, not this one. - INIT_MEMORY: "2G" - MAX_MEMORY: "5G" + # 2G heap inside a 3.2g container cap, for the 2 vCPU / 4 GB box. A JVM + # needs meaningful memory outside its heap: metaspace, code cache, GC + # structures, thread stacks and direct byte buffers. Setting the heap + # equal to the container limit is the classic way to get the kernel + # OOM-killer to kill the server under load, which looks exactly like a + # random crash. ~1.2g of headroom is the margin that prevents it. + # The host has 4 GB total, so a 3.2g container leaves ~800Mi for the + # kernel, Docker, sshd, tailscaled, auditd and the deploy timer. + # + # BE HONEST ABOUT THIS NUMBER: 2G is under-provisioned for the pack this + # server actually runs - roughly 30 mods including Create, CC:Tweaked, + # BlueMap, Terralith, When Dungeons Arise and Ad Astra's five extra + # dimensions. It ran on 5G before. Expect GC pressure and low TPS, and + # if it will not hold, the fix is to cut load rather than to raise this + # above the container cap: BlueMap's renderer is the heaviest consumer + # that costs no gameplay, then Ad Astra's dimension count, then view + # distance below. + INIT_MEMORY: "1G" + MAX_MEMORY: "2G" # PrismProtect's sqlite-jdbc can't locate its own bundled native lib # under Forge's per-mod module classloader and throws # NativeLibraryNotFoundException, leaving its DB connection null and @@ -250,10 +266,22 @@ services: # If this ever opens to players outside the family, revisit it. ALLOW_FLIGHT: "true" - MAX_PLAYERS: "20" + # 5, not 20. 20 was typed, never measured, and never tested: the + # memory and tick cost of a slot is only paid when someone is in it, + # but the WORST case is what has to fit in the box, and on 4 GB it + # does not. Five is the actual population (known family members), so + # this caps the worst case at something this plan can hold instead of + # at something it cannot. + MAX_PLAYERS: "5" ONLINE_MODE: "TRUE" # Java accounts only (blocks cracked/anonymous clients) - VIEW_DISTANCE: "12" - SIMULATION_DISTANCE: "8" + # Cut with the 2026-10-11 downsize (was 12 and 8). Loaded chunks scale + # with the square of view distance, so 12 -> 8 is about 56% fewer chunks + # held per player: the single biggest lever on both heap and tick time + # that costs no content. Simulation distance 6 still comfortably exceeds + # the 4-chunk radius that keeps a player's own machines and farms + # ticking, so Create contraptions near a player keep running. + VIEW_DISTANCE: "8" + SIMULATION_DISTANCE: "6" TZ: "America/New_York" volumes: - ./data:/data