KV get fails to get republished value (message) with multiple headers

[a-mytra]

What version were you using?

nats-server: v2.9.21 nats.ws 1.16.1 client running on Chrome browser

What environment was the server running in?

PRETTY_NAME="Debian GNU/Linux 11 (bullseye)" NAME="Debian GNU/Linux" VERSION_ID="11" VERSION="11 (bullseye)" VERSION_CODENAME=bullseye ID=debian HOME_URL="https://www.debian.org/" SUPPORT_URL="https://www.debian.org/support" BUG_REPORT_URL="https://bugs.debian

Is this defect reproducible?

Steps to reproduce

I create a KV named B

nats kv add B

I create a stream named DC_B to subscribe to all message on the subject A.> and republish to $KV.B.>

nats stream add DC_B 
      --subjects='A.>'
      --republish-source='>'
      --republish-destination='$KV.B.>'
      --storage=memory 
      --replicas=1 
      --retention=limits 
      --discard=old 
      --max-msgs=1 
      --max-msgs-per-subject=-1 
      --max-bytes=-1 
      --max-age=-1 
      --max-msg-size=-1 
      --dupe-window="2m0s" 
      --allow-rollup 
      --no-deny-delete 
      --no-deny-purge

I publish a message on A

nats pub "A.tomato" "hello"

As a sanity check I also explicitly put a value directly into B to compare

nats kv put B 'A.potato' "yo"

It does get capture by the stream DC_B and does get republished to the KV bucket B correctly

-> A.tomato "hello"
    -> KV B
        -> A.tomato: "hello"

Watching with nats sub '>' while this is happening I can see the message that goes to B contains multiple headers

[#48] Received on "_INBOX.Olz80ftB4Yj0cBmlJhobQO.XOE2JA36"
Nats-Stream: DC_B
Nats-Stream: KV_B
Nats-Subject: A.tomato                 <------ this
Nats-Subject: $KV.B.A.tomato           <------ and this
Nats-Sequence: 2
Nats-Sequence: 2
Nats-Last-Sequence: 1
Nats-Time-Stamp: 2023-09-02T14:48:06.891798Z

hello

[#43] Received on "_INBOX.qppVDoINpVj2TabJ60qInw.LhKzBUHR"
Nats-Stream: KV_B
Nats-Subject: $KV.B.A.potato
Nats-Sequence: 3
Nats-Time-Stamp: 2023-09-02T14:52:48.235748Z

yo

Using the CLI I can get the value using the key A.tomato from B as expected 👍

nats kv get B "A.tomato"

hello

This also works using the Python client correctly. 👍

Here is the bug 🐛 in the JS implementation 👇

const B = await js.views.kv("B")
let tomato = await B.get("A.tomato")
let potato = await B.get("A.potato")

I get tomato is null 😞 and and potato is correctly "yo"

The only difference is that A.tomato was republished into the KV and has multiple headers.

I tracked it down to the Bucket class's get function here https://github.com/nats-io/nats.deno/blob/db885246875d3c1d6401e1f686a9a0570ac20e9b/jetstream/kv.ts#L528 This equality fails

      if (ke.key !== ek) {

And it fails due to the incorrect header retrieval in thelength getter in the KvStoredEntryImpl class here https://github.com/nats-io/nats.deno/blob/db885246875d3c1d6401e1f686a9a0570ac20e9b/jetstream/kv.ts#L984C5

get length() {
    const slen = this.sm.header.get(JsHeaders.MessageSizeHdr) || ""
    if (slen !== "") {
        return parseInt(slen, 10)
    }
    return this.sm.data.length
}

The call to this.sm.header.get randomly (?) returns one of the headers for Nats-Subject so it doesn't match the subject equality check.

The value is correctly being retrieved but it is rejected and thrown out.

Given the capability you are leveraging, describe your expectation?

I expect

const B = await js.views.kv("B")
let tomato = await B.get("A.tomato")

to return the value stored.

Given the expectation, what is the defect you are observing?

Multiple headers (specifically Nats-Subject) in a KV key messages cause the value to be irretrievable by the JS implementation while Python and CLI (assuming Go as well) work.

[a-mytra]

For a quick fix on my end, I inject a new get function into a KV that the erroneous subject check

// --- INJECT FIX for broken multi-header bug
const AsyncFunction = async function () { }.constructor
const ngc = AsyncFunction("k", "opts", `
  const ek = this.encodeKey(k);
  this.validateKey(ek);
  let arg = {
      last_by_subj: this.subjectForKey(ek)
  };
  if (opts && opts.revision > 0) {
      arg = {
          seq: opts.revision
      };
  }
  let sm;
  try {
      if (this.direct) {
          const direct = this.jsm.direct;
          sm = await direct.getMessage(this.bucketName(), arg);
      } else {
          sm = await this.jsm.streams.getMessage(this.bucketName(), arg);
      }
      const ke = this.smToEntry(sm);
      return ke;
  } catch (err) {
      if (err.code === ErrorCode.JetStream404NoMessages) {
          return null;
      }
      throw err;
  }`)

// @ts-ignore
kv.get = ngc

I think a better implementation is to either use the correct headers for checking or drop the check altogether.

[aricart]

@a-mytra not sure if there's some interaction with the nats cli or the other tooling you are using. But if I run this test:

  const jsm = await nc.jetstreamManager()

  await jsm.streams.add({
    name: "A",
    subjects: ["A.>"],
    storage: StorageType.Memory,
    republish: {
      src: ">",
      dest: "$KV.B.>"
    }
  });

  nc.subscribe("$KV.B.>", {callback: (_, msg) => {
    console.log(`${msg.subject} ${msg.headers ? msg.headers : "no headers"} ${msg.string()}`);
    }})

  const js = nc.jetstream();
  const B = await js.views.kv("B");

  nc.publish("A.orange", "hey");
  await js.publish("A.tomato", "hello");

  await B.put("A.potato", "yo");

  await nc.flush();

  const iter = await B.keys();
  for await(const v of iter) {
    console.log(">", v)
  }

$KV.B.A.orange NATS/1.0
Nats-Stream: A
Nats-Subject: A.orange
Nats-Sequence: 1
Nats-Last-Sequence: 0

 hey
$KV.B.A.tomato NATS/1.0
Nats-Stream: A
Nats-Subject: A.tomato
Nats-Sequence: 2
Nats-Last-Sequence: 0

 hello
$KV.B.A.potato no headers yo
> A.orange
> A.tomato
> A.potato

I do not get any duplicate headers, and your usecase works correctly. I tried this with your server version 2.9.21 and with the just released 2.10.1

[aricart]

Also what it looks like is that the KV is doing republishing. Can you print the kv configuration

[aricart]

OK take it back - I can reproduce the issue

[a-mytra]

Ok I'm glad, I was about to carve out some time to sit down and asciinema on Mac and Linux to demonstrate. I already pulled too much hair out and a couple of weekends figuring out what was going on.

The way I found out what was going on was by manually inspecting the binary frame on the websocket message reply after a .get and I realized we are indeed getting the value back from NATS but somehow the JS implementation was discarding it.

Thank you for looking into this 🙇

[jnmoyne]

Leaving the specific issue aside and stepping back a little bit I would suggest you upgrade to nats-server V 2.10.1 (and nats V 0.1.1) and simply use KV bucket sourcing rather than trying to feed the bucket from republished message.

e.g.

nats kv add A
nats kv add B --source A
nats kv put A a hello
nats kv get B a

Besides this being cleaner and simpler, it is also much better in terms of reliability and quality of service: with feeding the KV bucket from (Core NATS) republished message if for some reason or another the republished message gets dropped (maybe one of the nats-servers in the cluster getting restarted right at the same time for example) then that message would never make it to the destination KV bucket. Using sourcing between the buckets is safe from any Core NATS message getting dropped or any other issue with the server infrastructure and sourcing/mirroring between streams or KV is 'store and forward reliable'.

[aricart]

@a-mytra the issue is not a client issue - I believe the reason why it works for you with a different client has to do with the order the values come out. So that is will be specific to the language/map etc that is used on read. I suspect this is possibly a server issue, because I wouldn't expect Nats- headers to be persisted. On the direct gets for KV, the headers are added, and that is why we get messages with 2 values when we should only be getting one for those. Keep tuned

[aricart]

Thank you for looking into this 🙇

Actually thank you for filling the detailed issue.

[aricart]

I have a workaround for you, if you set the option on the javascript kv allow_direct: false you'll work around the issue

[aricart]

https://github.com/nats-io/nats-server/issues/4573