Edge Computing ও Cloudflare Workers — Isolation Spectrum-এর শেষ বিন্দু

লক্ষ্য: শুধু “isolate দ্রুত” বলার বদলে, একটা concrete multi-tenant churn-scenario-এ সেই গতির throughput/latency-প্রভাব সংখ্যায় দেখানো — ঠিক containers-vs-vms লেসনের IsolationBoundaryCalculator-এর মতো পদ্ধতিতে, এবার চারটা primitive নিয়ে।

# isolation_churn_sim.py -- models per-tenant startup cost for the FOUR
# isolation primitives this module covered, under a high-churn multi-tenant
# edge-style workload (many different tenants' code, requested often enough
# that warm-pool reuse cannot be assumed).
#
# Every constant is either a cited, published order-of-magnitude figure or
# clearly labeled as this lesson's own illustrative estimate:
#   full_vm       ~30s    -- general VM boot; Felter et al., IEEE ISPASS 2015,
#                            and the containers-vs-vms lesson's cited 30-60s
#                            range for a typical cloud VM boot
#   microvm       ~125ms  -- Firecracker microVM boot; Agache et al.,
#                            USENIX NSDI 2020 (measured by AWS for Lambda)
#   container     ~50ms   -- OS-level container start INCLUDING app runtime
#                            init (not just the bare clone() syscall, which
#                            is sub-millisecond) -- order-of-magnitude figure
#                            consistent with Felter et al. 2015
#   v8_isolate    ~5ms    -- Cloudflare Workers V8 isolate cold start;
#                            Cloudflare Blog, "Eliminating Cold Starts with
#                            Cloudflare Workers" (2018)

COLD_START_SECONDS = {
    "full_vm":    30.0,
    "microvm":     0.125,
    "container":   0.05,
    "v8_isolate":  0.005,
}
LABELS = {
    "full_vm":    "Full VM (BIOS/UEFI + kernel + init)",
    "microvm":    "microVM (Firecracker)",
    "container":  "Container (OS-level, incl. app runtime init)",
    "v8_isolate": "V8 isolate (Cloudflare Workers)",
}

# ---- (1) total cold-start cost across a high-churn request volume ----
# Worst case: every one of N requests needs a FRESH isolation instance.
N_REQUESTS = 1_000_000

print("=== total cold-start cost, 1,000,000 fresh per-tenant starts ===")
for key in ("full_vm", "microvm", "container", "v8_isolate"):
    total_s = COLD_START_SECONDS[key] * N_REQUESTS
    print(f"{LABELS[key]:<48} : {total_s:>12,.1f} s  = {total_s / 3600:>9,.2f} hours")
print()

# ---- (2) Little's Law: concurrent "starter" capacity needed to sustain ----
# a target new-tenant-start rate (this lesson's own illustrative scenario).
# L = lambda * W  (concurrency = arrival rate * time-in-system)
TARGET_STARTS_PER_SEC = 5_000

print(f"=== concurrent 'in-flight start' capacity needed @ {TARGET_STARTS_PER_SEC:,}/s new-tenant churn ===")
concurrency = {}
for key in ("full_vm", "microvm", "container", "v8_isolate"):
    c = TARGET_STARTS_PER_SEC * COLD_START_SECONDS[key]
    concurrency[key] = c
    print(f"{LABELS[key]:<48} : {c:>10,.1f} concurrent starts in flight")
print()

print("=== ratio vs. the isolate model ===")
base = concurrency["v8_isolate"]
for key in ("full_vm", "microvm", "container"):
    print(f"{LABELS[key]:<48} : {concurrency[key] / base:>8,.0f}x more concurrent start-capacity needed than isolates")

প্রকৃত চালিয়ে পাওয়া output (deterministic — হুবহু এটাই আসবে):

=== total cold-start cost, 1,000,000 fresh per-tenant starts ===
Full VM (BIOS/UEFI + kernel + init)              : 30,000,000.0 s  =  8,333.33 hours
microVM (Firecracker)                            :    125,000.0 s  =     34.72 hours
Container (OS-level, incl. app runtime init)     :     50,000.0 s  =     13.89 hours
V8 isolate (Cloudflare Workers)                  :      5,000.0 s  =      1.39 hours

=== concurrent 'in-flight start' capacity needed @ 5,000/s new-tenant churn ===
Full VM (BIOS/UEFI + kernel + init)              :  150,000.0 concurrent starts in flight
microVM (Firecracker)                            :      625.0 concurrent starts in flight
Container (OS-level, incl. app runtime init)     :      250.0 concurrent starts in flight
V8 isolate (Cloudflare Workers)                  :       25.0 concurrent starts in flight

=== ratio vs. the isolate model ===
Full VM (BIOS/UEFI + kernel + init)              :    6,000x more concurrent start-capacity needed than isolates
microVM (Firecracker)                            :       25x more concurrent start-capacity needed than isolates
Container (OS-level, incl. app runtime init)     :       10x more concurrent start-capacity needed than isolates

দুইটা ফলাফল একসাথে পড়লে high-churn edge workload-এ কেন isolate মডেল ব্যবহারিকভাবে জেতে তা স্পষ্ট হয়। প্রতি-সেকেন্ড ৫,০০০ নতুন-tenant churn সামলাতে full VM মডেলে ১,৫০,০০০টা concurrent VM-বুট একসাথে চালাতে হবে — বাস্তবে অসম্ভব একটা সংখ্যা। microVM-এও ৬২৫টা concurrent boot দরকার। Container নামিয়ে আনে ২৫০-এ। V8 isolate মাত্র ২৫টা concurrent “start”-এই সেই একই churn সামলায় — full VM-এর তুলনায় ৬,০০০ গুণ কম resource। এটাই precisely সেই সংখ্যাগত কারণ কেন edge platform-এর architecture isolate-ভিত্তিক না হয়ে উপায় নেই।