ai2070-net 0.9.0

High-performance, schema-agnostic event bus for AI runtime workloads

Net Python

High-performance, schema-agnostic event bus for AI runtime workloads.

Installation

pip install ai2070-net

The package publishes as ai2070-net on PyPI but imports as from net import ... (the in-source module name is preserved). For the higher-level Pythonic surface (generators, typed channels, dataclass/Pydantic support), install ai2070-net-sdk instead — it depends on this package.

Quick Start

from net import Net

# Create event bus (defaults to CPU core count shards)
bus = Net()

# Ingest events - fast path with raw JSON strings (23M+ ops/sec)
bus.ingest_raw('{"token": "hello", "index": 0}')

# Or use dict for convenience (4M+ ops/sec)
bus.ingest({"token": "world", "index": 1})

# Batch ingestion for maximum throughput
events = [f'{{"token": "tok_{i}"}}' for i in range(10000)]
count = bus.ingest_raw_batch(events)

# Poll events
response = bus.poll(limit=100)
for event in response:
    print(event.raw)
    # Or parse to dict
    data = event.parse()

# Check stats
stats = bus.stats()
print(f"Ingested: {stats.events_ingested}, Dropped: {stats.events_dropped}")

# Shutdown
bus.shutdown()

Context Manager

with Net(num_shards=4) as bus:
    bus.ingest_raw('{"data": "value"}')
# Automatically shuts down

Configuration

bus = Net(
    num_shards=8,                    # Number of parallel shards
    ring_buffer_capacity=1_048_576,  # Events per shard (must be power of 2)
    backpressure_mode="drop_oldest", # What to do when full
)

Net Encrypted UDP Transport

Net provides encrypted point-to-point UDP transport for high-performance scenarios:

from net import Net, generate_net_keypair
import os

# Generate keypair for responder
keypair = generate_net_keypair()
psk = os.urandom(32).hex()

# Responder side
responder = Net(
    num_shards=2,
    net_bind_addr='127.0.0.1:9001',
    net_peer_addr='127.0.0.1:9000',
    net_psk=psk,
    net_role='responder',
    net_secret_key=keypair.secret_key,
    net_public_key=keypair.public_key,
    net_reliability='light',  # 'none', 'light', or 'full'
)

# Initiator side (knows responder's public key)
initiator = Net(
    num_shards=2,
    net_bind_addr='127.0.0.1:9000',
    net_peer_addr='127.0.0.1:9001',
    net_psk=psk,
    net_role='initiator',
    net_peer_public_key=keypair.public_key,
)

# Use as normal
initiator.ingest_raw('{"event": "data"}')

Backpressure Modes

• "drop_newest" - Reject new events when buffer is full
• "drop_oldest" - Evict oldest events to make room
• "fail_producer" - Raise an error

NAT traversal (optimization, not correctness)

Two NATed peers already reach each other through the mesh's routed-handshake path. NAT traversal opens a shorter direct path when the NAT shape allows it; it's never required for connectivity. Every method below is safe to call regardless of NAT type — a failed punch or a traversal: * RuntimeError is not a connectivity failure, traffic keeps riding the relay. The whole surface is a no-op when the native module was built without --features nat-traversal: every call raises RuntimeError("traversal: unsupported").

from net import NetMesh

mesh = NetMesh(bind_addr="0.0.0.0:9000", psk="00" * 32)

mesh.reclassify_nat()

klass  = mesh.nat_type()            # "open" | "cone" | "symmetric" | "unknown"
reflex = mesh.reflex_addr()         # "203.0.113.5:9001" or None

observed = mesh.probe_reflex(peer_node_id)   # "ip:port"

# Attempt a direct connection via the pair-type matrix.
# `coordinator` mediates the punch when the matrix picks one.
# Always returns — inspect stats to learn which path won.
mesh.connect_direct(peer_node_id, peer_pubkey_hex, coordinator_node_id)

# Cumulative counters — all int, monotonic.
s = mesh.traversal_stats()
s.punches_attempted   # coordinator mediated a PunchRequest + Introduce
s.punches_succeeded   # ack arrived AND direct handshake landed
s.relay_fallbacks     # landed on the routed path after skip/fail

Operators with a known-public address skip the classifier sweep entirely. The override pins "open" + the supplied address on every capability announcement; call announce_capabilities() after to propagate (the setter resets the rate-limit floor so the next announce is guaranteed to broadcast).

mesh.set_reflex_override('203.0.113.5:9001')
mesh.announce_capabilities(caps)
# later:
mesh.clear_reflex_override()
mesh.announce_capabilities(caps)

Traversal failures surface as RuntimeError with a stable traversal: <kind>[: <detail>] message prefix. The <kind> discriminator is one of reflex-timeout | peer-not-reachable | transport | rendezvous-no-relay | rendezvous-rejected | punch-failed | port-map-unavailable | unsupported. Match on the prefix for machine-readable branching:

try:
    mesh.connect_direct(peer_node_id, peer_pubkey_hex, coord_id)
except RuntimeError as e:
    msg = str(e)
    if msg.startswith("traversal: unsupported"):
        ...   # native module built without --features nat-traversal
    elif msg.startswith("traversal: peer-not-reachable"):
        ...

"unsupported" is the signal that the bindings are linked unconditionally and the native module doesn't have the feature — callers can branch cleanly without probing for symbol presence.

Channels (distributed pub/sub)

Named pub/sub over the encrypted mesh. Publishers register channels with access policy; subscribers ask to join via a membership subprotocol; publish fans payloads out to every current subscriber.

from net import NetMesh, ChannelAuthError, ChannelError

pub = NetMesh('127.0.0.1:9001', '42' * 32)
try:
    pub.register_channel(
        'sensors/temp',
        visibility='global',      # or 'subnet-local' | 'parent-visible' | 'exported'
        reliable=True,
        priority=2,
        max_rate_pps=1000,
    )

    # Subscriber side (after handshake with pub):
    # sub.subscribe_channel(pub.node_id, 'sensors/temp')

    # Fan a payload out to all subscribers.
    report = pub.publish(
        'sensors/temp',
        b'{"celsius": 22.5}',
        reliability='reliable',
        on_failure='best_effort',
        max_inflight=32,
    )
    print(f"{report['delivered']}/{report['attempted']} subscribers received")
finally:
    pub.shutdown()

# Typed errors for ACL outcomes:
# try: sub.subscribe_channel(peer_id, 'restricted')
# except ChannelAuthError: ...   # publisher denied
# except ChannelError: ...       # unknown channel / other rejection

Channel names always cross the binding as strings (not the u16 hash) to avoid ACL bypass via collision. The Python binding does not yet expose a dedicated per-channel receive API; that is a follow-up.

CortEX & NetDb (event-sourced state)

Typed, event-sourced state on top of RedEX — tasks and memories with filterable queries and sync watch iterators. Includes the snapshot_and_watch primitive whose race fix landed on v2, so you can safely "paint what's there now, then react to changes" without losing updates that race during construction.

from net import NetDb, CortexError

db = NetDb.open(origin_hash=0xABCDEF01, with_tasks=True, with_memories=True)
tasks = db.tasks

try:
    seq = tasks.create(1, 'write docs', 100)
    tasks.wait_for_seq(seq)   # block until the fold has applied
except CortexError as e:
    # adapter-level failure (RedEX I/O, fold halted, etc.)
    ...

# Snapshot + watch, one atomic call — no race.
snap, it = tasks.snapshot_and_watch_tasks(status='pending')
print('initial:', len(snap), 'pending tasks')
for batch in it:
    print('update:', len(batch), 'pending tasks')
    if len(batch) == 0:
        it.close()    # idempotent; ends the iterator
        break

db.close()

Standalone adapters

If you only need one model, skip the NetDb facade:

from net import Redex, TasksAdapter

redex = Redex(persistent_dir='/var/lib/net/redex')
tasks = TasksAdapter.open(redex, origin_hash=0xABCDEF01, persistent=True)

MemoriesAdapter exposes the same shape with store / retag / pin / unpin / delete / list_memories / watch_memories / snapshot_and_watch_memories.

Raw RedEX file (no CortEX fold)

For domain-agnostic persistent logs — your own event schema, no fold, no typed adapter — open a RedexFile directly from a Redex. The tail is a sync Python iterator; call close() or let StopIteration fire when the file closes.

from net import Redex, RedexError

redex = Redex(persistent_dir='/var/lib/net/events')
file = redex.open_file(
    'analytics/clicks',
    persistent=True,
    fsync_interval_ms=100,           # or fsync_every_n=1000
    retention_max_events=1_000_000,
)

# Append (or batch-append).
seq = file.append(b'{"url": "/home"}')
# `append_batch` returns the first-seq int of the batch, or `None`
# for an empty input. The `None` return is the explicit "I
# appended nothing" signal — pre-`bugfixes-8` it returned `0`,
# which collided with the legitimate "first event of a non-empty
# batch landed at seq 0" return.
first = file.append_batch([b'{"a": 1}', b'{"a": 2}'])

# Tail — backfills the retained range, then streams live appends.
try:
    for event in file.tail(from_seq=0):
        print(event.seq, bytes(event.payload))
        if should_stop:
            break           # idempotent; ends the iterator via close()
except RedexError as e:
    ...

file.close()

Errors from the RedEX surface raise RedexError (invalid channel name, bad config, append / tail / sync / close failures).

Why snapshot_and_watch_*?

Calling list_tasks() then watch_tasks() takes two independent state reads. A mutation landing between them would be silently lost under the old skip(1) implementation. The atomic primitive returns the snapshot and an iterator seeded so that any divergent initial emission is forwarded through instead of dropped — see docs/STORAGE_AND_CORTEX.md.

Redis Streams consumer-side dedup helper

The Net Redis adapter writes a stable dedup_id field on every XADD entry: {producer_nonce:hex}:{shard_id}:{sequence_start}:{i}. Combined with the bus's persistent producer-nonce path (producer_nonce_path on EventBusConfig), the id is stable across both within-process retries AND cross-process restart — the MULTI/EXEC-timeout race becomes filterable at consume time.

RedisStreamDedup is the consumer-side helper, exposed on the net PyO3 module:

from net import RedisStreamDedup
import redis

# ~10k events/sec * 1 min dedup window → ~600,000.
dedup = RedisStreamDedup(capacity=600_000)

r = redis.Redis(host="localhost", port=6379)
cursor = "0"
while True:
    # XRANGE bounds are INCLUSIVE on both ends. After the first
    # page we must use the exclusive form `(<id>` so we don't
    # re-read the entry the cursor points at — a vanilla
    # `min=cursor` loop spins forever once the cursor reaches the
    # tail and the same entry is returned every iteration.
    start = cursor if cursor == "0" else f"({cursor}"
    entries = r.xrange("net:shard:0", min=start, max="+", count=100)
    for entry_id, fields in entries:
        dedup_id = fields.get(b"dedup_id", b"").decode()
        if not dedup_id:
            # No dedup_id → older entry or non-Net producer; skip
            # dedup and process as-is.
            process(entry_id, fields)
            continue
        if not dedup.is_duplicate(dedup_id):
            process(entry_id, fields)
        cursor = entry_id.decode()
    if not entries:
        break

Surface:

dedup = RedisStreamDedup()                # default capacity 4096
dedup = RedisStreamDedup(capacity=N)      # explicit; 0 → 1
dedup.is_duplicate(id: str) -> bool       # test-and-insert
dedup.len                                 # property — tracked-id count
dedup.capacity                            # property — configured cap
dedup.is_empty                            # property
dedup.clear()                             # reset (e.g. on consumer-group rebalance)

The helper is transport-agnostic — bring your own redis-py / aioredis / equivalent client; it just answers the dedup question against an in-memory LRU. Concurrency: each handle wraps a Rust Mutex<RedisStreamDedup>, so concurrent calls from multiple Python threads are safe but serialize. Production-shape is one helper per consumer thread.

Security Surface (Stage A–E)

The mesh layer surfaces the same identity / capabilities / subnets / channel-auth story that the Rust SDK and the TypeScript / Node SDKs ship. Full staging and rationale: docs/SDK_SECURITY_SURFACE_PLAN.md. Python-binding parity details: docs/SDK_PYTHON_PARITY_PLAN.md.

Identity + permission tokens

Every node has an ed25519 identity; permission tokens are ed25519- signed delegations that authorize a subject to publish / subscribe / delegate / admin on a channel, optionally with further delegation depth.

from net import Identity, parse_token, verify_token, delegate_token

alice = Identity.generate()
bob = Identity.generate()

# Alice issues Bob a subscribe+delegate token good for 5 min, with
# one re-delegation hop remaining. `ttl_seconds=0` raises
# `TokenError` — a zero TTL would mint a born-expired token that
# every receiver would reject as `Expired`, leaving the issuer to
# diagnose the misuse from receiver-side log lines.
token = alice.issue_token(
    subject=bob.entity_id,
    scope=["subscribe", "delegate"],
    channel="sensors/temp",
    ttl_seconds=300,
    delegation_depth=1,
)
assert verify_token(token) is True

# Bob re-delegates to Carol; depth drops to 0 (leaf).
carol = Identity.generate()
child = delegate_token(bob, token, carol.entity_id, ["subscribe"])
assert parse_token(child)["delegation_depth"] == 0

Capability announcements + peer discovery

Announce hardware / software / model / tool / tag fingerprints, then query the local capability index with a filter.

mesh.announce_capabilities({
    "hardware": {
        "cpu_cores": 16,
        "memory_mb": 65536,
        "gpu": {"vendor": "nvidia", "model": "h100", "vram_mb": 81920},
    },
    "models": [{"model_id": "llama-3.1-70b", "family": "llama",
                "context_length": 128_000}],
    "tags": ["gpu", "prod"],
})

gpu_peers = mesh.find_nodes({
    "require_gpu": True,
    "gpu_vendor": "nvidia",
    "min_vram_mb": 40_000,
})

Scoped discovery (reserved scope:* tags)

A provider can narrow who its query result reaches by tagging its CapabilitySet with reserved scope:* tags. Queries call mesh.find_nodes_scoped(filter, scope) to filter candidates. The wire format and forwarders are untouched — enforcement is purely query-side.

# GPU pool advertised to one tenant only.
mesh.announce_capabilities({
    "tags": ["model:llama3-70b", "scope:tenant:oem-123"],
})

# Tenant-scoped query — returns this node + any Global (untagged) peers.
oem_nodes = mesh.find_nodes_scoped(
    {"require_tags": ["model:llama3-70b"]},
    {"kind": "tenant", "tenant": "oem-123"},
)

Accepted scope dict shapes: {"kind": "any"} (default), {"kind": "global_only"}, {"kind": "same_subnet"}, {"kind": "tenant", "tenant": "<id>"}, {"kind": "tenants", "tenants": [...]}, {"kind": "region", "region": "<name>"}, {"kind": "regions", "regions": [...]}. Reserved announcement tags: scope:subnet-local (visible only under same_subnet), scope:tenant:<id>, scope:region:<name> — strictest scope wins. Untagged peers resolve to Global and stay visible under permissive queries. Full design: docs/SCOPED_CAPABILITIES_PLAN.md.

Capability propagation is multi-hop, bounded by MAX_CAPABILITY_HOPS = 16 with (origin, version) dedup on every forwarder. capability_gc_interval_ms controls both the index TTL sweep and the dedup cache eviction. See docs/MULTIHOP_CAPABILITY_PLAN.md.

Subnets

Nodes can bind to a hierarchical SubnetId (1–4 levels, each 0–255) directly, or derive one from announced tags via a SubnetPolicy.

# Explicit subnet.
mesh = NetMesh("127.0.0.1:9000", PSK, subnet=[3, 7, 2])

# Or derive from tags.
mesh = NetMesh(
    "127.0.0.1:9001", PSK,
    subnet_policy={
        "rules": [
            {"tag_prefix": "region:", "level": 0,
             "values": {"eu": 1, "us": 2, "apac": 3}},
            {"tag_prefix": "zone:", "level": 1,
             "values": {"a": 1, "b": 2, "c": 3}},
        ]
    },
)

Channel authentication

Publishers set publish_caps / subscribe_caps / require_token on register_channel. Subscribers present a PermissionToken via the optional token=bytes kwarg on subscribe_channel.

mesh.register_channel(
    "gpu/jobs",
    subscribe_caps={"require_gpu": True, "min_vram_mb": 16_000},
    require_token=True,
)

# Subscriber side, with a token issued by the publisher:
mesh.subscribe_channel(publisher_node_id, "gpu/jobs", token=token_bytes)

Denied subscribes raise ChannelAuthError (a subclass of ChannelError); malformed tokens raise TokenError whose message has the form "token: <kind>" (invalid_signature, expired, delegation_exhausted, …). Successful subscribes populate an AuthGuard bloom filter on the publisher so every subsequent publish admits the subscriber in constant time. An expiry sweep (default 30 s) evicts subscribers whose tokens age out; a per- peer auth-failure rate limiter throttles bad-token storms. Cross- SDK behaviour is fixed by the Rust integration suite; see tests/channel_auth.rs and tests/channel_auth_hardening.rs.

Compute (daemons + migration)

Run MeshDaemons directly from Python. DaemonRuntime owns the factory table, the per-daemon hosts, and the Registering → Ready → ShuttingDown lifecycle gate that decides when inbound migrations may land. Daemons are any Python object whose process(event) returns a list of bytes/bytearray payloads — the runtime wraps each output in a causal link and forwards it.

Build the native module with the compute feature (maturin picks it up on the default build) and import from net. Full design notes: docs/SDK_COMPUTE_SURFACE_PLAN.md.

from net import DaemonRuntime, NetMesh, Identity, CausalEvent


class EchoDaemon:
    """Stateless echo — ships every event's payload straight back."""

    name = "echo"

    def process(self, event: CausalEvent) -> list[bytes]:
        return [bytes(event.payload)]

    # Optional: snapshot() / restore(state) for migration-capable daemons.


mesh = NetMesh("127.0.0.1:9000", "42" * 32)
rt = DaemonRuntime(mesh)

# 1. Register factories BEFORE flipping the runtime to Ready.
rt.register_factory("echo", lambda: EchoDaemon())

# 2. Ready the runtime — after this point spawn / migration accept.
rt.start()

# 3. Spawn a daemon; Identity pins the ed25519 keypair so
#    origin_hash / entity_id stay stable across migrations.
identity = Identity.generate()
handle = rt.spawn("echo", identity)
print(f"origin = 0x{handle.origin_hash:08x}")

# 4. Manually feed an event for testing; real delivery happens
#    via the mesh's causal chain.
event = CausalEvent(handle.origin_hash, sequence=1, payload=b"hello")
outputs = rt.deliver(handle.origin_hash, event)

# 5. Clean shutdown.
rt.stop(handle.origin_hash)
rt.shutdown()

process must be synchronous — the core's contract is sync, and the PyO3 bridge re-attaches the GIL for the duration of each call. Raising inside process surfaces as DaemonError on the caller.

Migration

start_migration(origin_hash, source_node, target_node) orchestrates the six-phase cutover (Snapshot → Transfer → Restore → Replay → Cutover → Complete). The source seals the daemon's ed25519 seed into the outbound snapshot using the target's X25519 static pubkey; the target rebuilds the daemon via the factory registered under the same kind, replays any events that arrived during transfer, then activates.

from net import MigrationError, migration_error_kind

try:
    mig = rt.start_migration(handle.origin_hash, src_node_id, dst_node_id)
    # mig.phase  — "snapshot" | "transfer" | "restore" | ...
    # mig.source_node / mig.target_node
    mig.wait()                      # blocks to completion
except MigrationError as e:
    kind = migration_error_kind(e)
    if kind == "not-ready":               ...  # target start() didn't run
    elif kind == "factory-not-found":     ...  # target missing this kind
    elif kind == "compute-not-supported": ...  # target has no DaemonRuntime
    elif kind == "state-failed":          ...  # snapshot / restore threw
    elif kind == "identity-transport-failed": ...  # seal / unseal failed
    # ...see SDK_COMPUTE_SURFACE_PLAN.md for the full enum

start_migration_with(origin, src, dst, opts) exposes options such as seal_seed=False for test scenarios. On the target node, call rt.register_migration_target_identity(kind, identity) before any migration of that kind lands; without it the runtime rejects sealed-seed envelopes with migration_error_kind == "identity-transport-failed".

Surface at a glance

Method	Description
DaemonRuntime(mesh)	Construct against an existing NetMesh
rt.register_factory(kind, fn)	Install a factory (before start())
rt.start() / rt.shutdown()	Flip the lifecycle gate
rt.spawn(kind, identity, config=None)	Spawn a local daemon
rt.spawn_from_snapshot(kind, identity, bytes, config=None)	Rehydrate
rt.stop(origin)	Stop a local daemon
rt.snapshot(origin)	Capture bytes for persistence / migration
rt.deliver(origin, event)	Feed an event (returns list[bytes])
rt.start_migration(origin, src, dst)	Orchestrate a live migration
rt.register_migration_target_identity(kind, id)	Pin unseal keypair on target for kind
handle.origin_hash / entity_id / stats()	Per-daemon identity + stats
DaemonError / MigrationError	Typed exceptions; migration_error_kind(e) parses e.kind

Compute Groups (Replica / Fork / Standby)

HA / scaling overlays on top of DaemonRuntime. Build the native module with the groups feature (implies compute) to expose ReplicaGroup, ForkGroup, StandbyGroup, and the GroupError exception class.

• ReplicaGroup — N interchangeable copies of a daemon. Deterministic identity from group_seed + index, so a replacement respawned on another node has a stable origin_hash. Load-balances inbound events across healthy members; auto-replaces on node failure.
• ForkGroup — N independent daemons forked from a common parent at fork_seq. Unique keypairs, shared ancestry via a verifiable ForkRecord.
• StandbyGroup — active-passive replication. One member processes events; standbys hold snapshots and catch up via sync_standbys(). On active failure the most-synced standby promotes and replays the events buffered since the last sync.

from net import (
    DaemonRuntime, ForkGroup, GroupError, ReplicaGroup, StandbyGroup,
    group_error_kind,
)

rt = DaemonRuntime(mesh)
rt.register_factory("counter", lambda: CounterDaemon())

# --- ReplicaGroup ----------------------------------------------------
replicas = ReplicaGroup.spawn(
    rt, "counter",
    replica_count=3,
    group_seed=bytes([0x11] * 32),
    lb_strategy="consistent-hash",   # or "round-robin" / "least-load"
                                     #    / "least-connections" / "random"
)
origin = replicas.route_event({"routing_key": "user:42"})
rt.deliver(origin, event)
replicas.scale_to(5)                 # grow
replicas.on_node_failure(failed_node_id)   # respawn elsewhere

# --- ForkGroup -------------------------------------------------------
forks = ForkGroup.fork(
    rt, "counter",
    parent_origin=0xABCDEF01,
    fork_seq=42,
    fork_count=3,
    lb_strategy="round-robin",
)
assert forks.verify_lineage()
for record in forks.fork_records():
    print(record["forked_origin"], record["fork_seq"])

# --- StandbyGroup ----------------------------------------------------
hot = StandbyGroup.spawn(
    rt, "counter",
    member_count=3,                  # 1 active + 2 standbys
    group_seed=bytes([0x77] * 32),
)
rt.deliver(hot.active_origin, event)
hot.sync_standbys()                  # periodic catchup
# On active-node failure:
# new_origin = hot.on_node_failure(failed_node_id)  # auto-promotes

Typed errors

Failures raise GroupError (a subclass of DaemonError). Use group_error_kind(e) to parse the discriminator from the Rust side's daemon: group: <kind>[: detail] message prefix:

from net import GroupError, group_error_kind

try:
    ReplicaGroup.spawn(rt, "never-registered",
                       replica_count=2, group_seed=bytes(32))
except GroupError as e:
    kind = group_error_kind(e)
    if kind == "not-ready":               ...  # runtime.start() didn't run
    elif kind == "factory-not-found":     ...  # kind wasn't registered
    elif kind == "no-healthy-member":     ...  # routed on an all-down group
    elif kind == "invalid-config":        ...  # e.g. replica_count == 0
    elif kind in ("placement-failed",
                  "registry-failed",
                  "daemon"):              ...  # core failure — read e args

Full staging, wire formats, and rationale: docs/SDK_GROUPS_SURFACE_PLAN.md. Core semantics live in the main README.md#daemons.

Performance Tips

1. Use ingest_raw() for maximum throughput - Pass pre-serialized JSON strings
2. Use ingest_raw_batch() for bulk operations - Reduces per-call overhead
3. Increase ring_buffer_capacity - Larger buffers handle bursts better
4. Match num_shards to CPU cores - Default is optimal for most cases

License

Apache-2.0