NATS
v2.14.8Networking & Messaging2026년 10월 9일
정확성과 성능에 초점을 맞춘 유지보수 릴리스입니다. 레거시 게이트웨이 응답 기능이 제거되고, 관련 클라이언트 인그레스 동작이 제한되었습니다.
조치 필요 (1)
breaking클라이언트 인그레스의
$GR.응답 접두사 예약클라이언트 인그레스에서 레거시
$GR.응답 접두사를 예약합니다. 클라이언트가 이 접두사를 사용하는 게이트웨이 응답 주제에 메시지를 게시하지 못하게 변경되었습니다.#8615에 해당합니다.
영향 확인 (2)
breaking레거시
$GR.응답 접두사 지원 제거서버가
v2.보다 이전 버전이면 적용됩니다.1. 2 서버가
v2.보다 이전 버전일 때 사용되던 레거시1. 2 $GR.응답 접두사를 보내고 처리하는 기능이 제거되었습니다.#8615에 해당합니다.breaking프록시 URL 없는 원격 프록시 옵션 거부
원격 프록시 옵션을 설정하면서 프록시 URL을 지정하지 않은 경우에 적용됩니다.
프록시 URL을 지정하지 않은 상태에서 원격 프록시 옵션을 설정하면 해당 옵션이 거부됩니다.
#8701에 해당합니다.
그 외 기록된 변경 53건 전체fixes 39 · value changes 11 · additions 3
fixes (39)
- - Concurrent account JWT updates no longer invalidate stream imports and interrupt message delivery (#8610)
- - Configuration reloads no longer incorrectly disconnect operator mode clients that send an
nkey(#8609) - -
Connzrequests selecting a connection bycidnow correctly apply account, user and MQTT client ID filters (#8613) - - Pipelined client operations are no longer processed after a connection is rejected for exceeding the account's
max_connectionslimit (#8614) - - Queue subscription deny permissions are now checked against the mapped delivery subject (#8616)
- - Outstanding reply permissions are now preserved when inherited account permissions are refreshed without changing the response policy (#8620, contributed by @0xasritha)
- - Subscription deny filters are now rebuilt atomically when inherited account permissions are refreshed, preventing a window in which denied messages could be delivered (#8622, contributed by @0xasritha)
- - Unquoted configuration values beginning with a number and a size or exponent suffix, such as
0K1abc, are now correctly parsed as strings (#8263, contributed by @koriyoshi2041) - - Account JWT updates now reject invalid subject mappings instead of reporting success, and failed replacements no longer remove existing mappings (#8627, contributed by @vulragrag-star)
- - Cached subscription results are now cleared when an already connected client sends another
CONNECT, preventing stale results from being reused after an account change (#8641, reported by @Tan-JunWei) - - Fixed several data races affecting service imports and exports during configuration reloads and account updates (#8675)
- - Per-account routes no longer double count inbound account message and byte statistics (#8677)
- - TLS certificate pin validation now checks the correct pins after a configuration reload (#8679)
- - Fast batch ping no longer starts a new batch incorrectly (#8714)
- - Auth callout rejections are now reported as authentication errors instead of account connection limit errors (#8688, contributed by @aniketatgithub)
- - Scoped
ProxyRequiredsettings are now respected in JWTs and auth callouts (#8694) - - Subscription interest notifications are no longer missed when a subscription is added or removed concurrently (#8703)
- - Fixed issues with auth callout
auth_usersbypass - - Fixed a PDH counter buffer overflow on Windows
- - Stream scale-down requests are now rejected when the target replica count has no configured account tier (#8608)
- - Stale per-message TTL entries are now removed when their messages have already been deleted, avoiding repeated expiry scans and excess CPU and memory usage (#8595, contributed by @digitalstraw)
- - The maximum deliveries exceeded advisory is now sent correctly when a consumer's
max_deliveris lowered below a message's existing delivery count (#8605, contributed by @zeke-lin) - - The consumer created time is now preserved after recovery (#8621)
- - Filestore compaction now respects the configured sync settings when replacing a message block (#8629)
- - Fixed a race during filestore compaction that could cause consumers to skip messages or lose redelivery tracking (#8623, contributed by @as-clearview)
- - Concurrent updates to R1 consumers no longer lose API responses (#8628)
- - Stream and consumer create and delete requests received during metalayer catchup now receive responses once catchup completes (#8643)
- - R1 consumer create requests no longer lose their responses when the consumer is deleted immediately afterwards (#8643)
- - Fixed a data race when handling consumer leader stepdown requests (#8647)
- - Fixed a memory leak affecting pull requests and direct gets after a configuration reload when the system account is explicitly configured in the
accountsblock (#8650, contributed by @1995parham) - - Fixed a panic when shutting down a clustered JetStream server immediately after startup (#8707, contributed by @clwluvw)
- - Idempotent stream create requests with sources no longer time out after leader changes (#8703)
- - Recovered ephemeral consumers with
inactive_thresholdset are now cleaned up properly after a server restart (#8732) - - Publish permissions are now checked against the transformed subject sent over the leafnode connection, avoiding the need to also allow the original stream import subject (#8530, contributed by @fbarth-ifr)
- - Fixed incorrect pending byte accounting after a partial compressed write that could stall a leafnode connection and prevent it from reconnecting (#8636)
- - Closed leafnode connections are now correctly removed from the account's leafnode list, reducing retained memory and garbage collection pressure (#8637)
- - Remote proxies no longer incorrectly warn that
nats://URLs are unsupported or lose the hub's initialINFOmessage (#8701) - - Config load no longer breaks QoS1/QoS2 delivery when
max_ack_pendingis unset (#8671) - - MQTT packet lengths are validated earlier against
max_payload(#8692)
value changes (11)
- - 1.26.9 (#8740)
- - Stalled source and mirror flow control is now reported in stream info and warning logs, making missing
$JS.imports or exports easier to diagnose (#8633)FC. > - - Stream and consumer placement now prefers peers that have caught up with the metalayer, reducing delays in API responses (#8643)
- - Improved the performance of deleting entries from and encoding AVL sequence sets (#8646)
- - Fast batch publish now reports backward sequence gaps as well as forward gaps (#8699)
- - Stream pending message calculations avoid unnecessary filestore cache access and handle deleted sequence ranges more efficiently (#8680, #8681)
- - Stream restore waits for the preferred node to become leader, avoiding an incorrect create response during clustered restore (#8690)
- - Memory backed streams avoid copying message headers and bodies twice when storing messages (#8652, contributed by @randomizedcoder)
- - Leafz now reports whether a leafnode connection uses WebSockets (#8683)
- - Partial packets are no longer copied in full on every read, reducing allocations and buffer copying (#8692)
- - Leafnode clients are now shed alongside regular clients during lame duck shutdown, avoiding a thundering herd (#8695)
additions (3)
- - TLS state is now available to auth callouts when leafnodes connect over WebSockets (#8686, contributed by @jurry-duty)
- - Remote proxies now support TLS with
https://proxy URLs and an optional proxytlsblock (#8701) - - Leafnode interest and request/reply across account imports (#8730, contributed by @edjeffreys)
NATS 스택에 추가
조치가 필요한 릴리스가 나왔을 때 주간 메일로 알려드립니다. 이번 릴리스의 브레이킹 체인지 같은 것들입니다.