aboutsummaryrefslogtreecommitdiff
path: root/doc
diff options
context:
space:
mode:
Diffstat (limited to 'doc')
-rw-r--r--doc/proposals/2022-01-add-leaf-endpoint90
-rw-r--r--doc/proposals/2022-01-author-reader-terminology41
-rw-r--r--doc/proposals/2022-01-domain-hint51
-rw-r--r--doc/proposals/2022-01-get-endpoints46
-rw-r--r--doc/proposals/2022-01-log-url26
-rw-r--r--doc/proposals/2022-01-tree-head-endpoints118
6 files changed, 372 insertions, 0 deletions
diff --git a/doc/proposals/2022-01-add-leaf-endpoint b/doc/proposals/2022-01-add-leaf-endpoint
new file mode 100644
index 0000000..3123e02
--- /dev/null
+++ b/doc/proposals/2022-01-add-leaf-endpoint
@@ -0,0 +1,90 @@
+Proposal: change add-leaf endpoint
+
+Background
+---
+Right now a log returns HTTP status 200 OK if it will "try" to merge a submitted
+leaf into its Merkle tree. A submitter should not assume that logging happened
+until they see an inclusion proof that leads up to a (co)signed tree head.
+
+If a submitted leaf does not show up in the log despite seeing HTTP status 200
+OK, the submitter must resubmit it. When a resubmission is required/expected is
+undefined.
+
+The reason for this "try" behavior is that log operations become much easier,
+especially in self-hosted environments that do not rely on managed databases.
+In other words, it is OK to just be "pretty sure" that a submitted leaf will be
+persisted and sequenced, and "100%" sure after sequencing actually happened.
+
+Proposal
+---
+A log should not return HTTP status 200 OK unless:
+1. The submitted leaf has been sequenced as part of a persisted database.
+2. The next tree head that the log signs will contain the submitted leaf.
+
+HTTP status 3XX is returned with, e.g., "Error=leaf has not been sequenced yet"
+if it is not guaranteed that the submitted leaf has been sequenced.
+
+This means that logging should be assumed after seeing HTTP status 200 OK. This
+assumption will be confirmed when the submitter obtains the next (co)signed tree
+head. Further investigation is required if it turns out that this assumption is
+false.
+
+Notes
+---
+An earlier draft of this proposal considered if useful debug information should
+be returned, such as "leaf index", "leaf hash", and "estimated time until a
+cosigned tree head is available". We decided to not go in this direction to
+avoid redundant and unsigned output that may be mis-used and tampered with ("not
+consistent with design").
+
+(Note that it is easy to determine when the next cosigned tree head will be
+available. The to-sign tree head has a timestamp, and it is rotated every 300s.
+Then it takes an additional 300s before the to-sign tree head is served with
+collected cosignatures.)
+
+An earlier draft of this proposal also considered to have verifiable output:
+ * Option 1: An inclusion proof and a signed tree head
+ * Option 2: An inclusion proof and a cosigned tree head
+
+This could be a worthwhile direction if the submitter can only obtain the
+required data by using the add-leaf endpoint, thus "forcing resubmits until the
+desired output is obtained". Credit to Al Cutter who proposed this (very nice)
+idea to us a while back.
+
+It is not appropriate to always return an inclusion proof for a signed tree
+head. What we want is for submitters to get inclusion proofs that reference
+cosigned tree heads.
+
+There are drawbacks to replace the above signed tree head with a cosigned tree
+head:
+ * A submitter that submits multiple leaves will likely (have to?) retrieve
+ the same cosigned tree head multiple times via the add-leaf endpoint. That
+ overhead adds up.
+ * A submitter will have to be in a "resubmit phase" for several minutes as
+ the default, because it takes time before a cosigned tree head becomes
+ available.
+ * (The most sensible implementation would likely resubmit periodically,
+ say, once per minute. A clever implementation would look at the
+ timestamp of the to-sign endpoint to determine when is the earliest time
+ that a merged may have happened.)
+
+Moreover, removing the get-inclusion-proof and get-tree-head-cosigned endpoints
+to force usage of add-leaf excludes (or makes for wonky) usage patterns of the
+log:
+ * "I just want to download all cosigned tree heads to archive them" -> add
+ leaves.
+ * "I just want to debug/know that the log is committed to have the leaf
+ logged, and rely on other witnesses" -> still forced to observe the log's
+ cosignatures.
+ * "I want an inclusion proof to a particular tree head" -> build the Merkle
+ tree yourself to construct that proof. The log's API chooses tree heads for
+ you.
+ * (Keeping these endpoints in addition to any new add-leaf output would to
+ some degree defeat the purpose of adding output, which is why it is not
+ considered an option.)
+
+In gist, we decided to go with a solution that is somewhere in between what we
+did before and what Al Cutter proposed. We defined when a resubmission is (not)
+expected. As a result, a self-hosted log may return at least one HTTP 3XX for
+each leaf request, and a few seconds later return HTTP status 200 OK for the
+same input data.
diff --git a/doc/proposals/2022-01-author-reader-terminology b/doc/proposals/2022-01-author-reader-terminology
new file mode 100644
index 0000000..fb447d2
--- /dev/null
+++ b/doc/proposals/2022-01-author-reader-terminology
@@ -0,0 +1,41 @@
+Start using the terminology "author" and "reader" proposal
+
+Background
+---
+Figure 1 in doc/design.md refers to
+
+ a) the party producing a signed checksum as "Signer", and
+ b) the party verifying a signed checksum as "Verifier".
+
+This is fine in isolation, but less appropriate when looking at it from a
+broader Sigsum perspective. For example, a "Signer" may also be a "Submitter".
+It seems like we are mixing terminology for roles and concrete actors here.
+
+The above is also ambiguous. For example, logs and witnesses sign things;
+witnesses and monitors verify things.
+
+Proposal
+---
+1) Replace "Signer" with "Author" when we are talking about a concrete party.
+
+According to Wikipedia's definition (https://en.wikipedia.org/wiki/Author), an
+'author is "the person who originated or gave existence to anything" and whose
+authorship determines responsibility for what was created'. This seems
+appropriate for us.
+
+The term "author" has been used in academic litterature before us for similar
+purposes:
+ * "In the setting of transparency logging [18] as depicted in Fig. 1, the
+ author generates events intended for recipients that describe data
+ processing by the author as it takes place"
+ * Link to paper: https://link.springer.com/chapter/10.1007/978-3-319-45741-3_7
+
+2) Replace "Verifier" with "Reader" when we are talking about a concrete party.
+
+According to Wikipedia's definition (https://en.wikipedia.org/wiki/Reading),
+"[r]eading is the process of taking in the sense or meaning of letters, symbols,
+etc., especially by sight or touch". Although the latter is not a perfect
+description for us, the first part is quite close and we could argue that we are
+in the "etc" category.
+
+The main idea here is that it should feel intuitive that an author has readers.
diff --git a/doc/proposals/2022-01-domain-hint b/doc/proposals/2022-01-domain-hint
new file mode 100644
index 0000000..322d9cc
--- /dev/null
+++ b/doc/proposals/2022-01-domain-hint
@@ -0,0 +1,51 @@
+Proposal: stricter domain hint requirements
+
+Background
+---
+Right now a log is expected to look up a submitter's public key hash via DNS. A
+domain hint, say, example.com, specifies the location of a TXT RR that contains
+the appropriate key hash in hex-encoding. "Some domain knows about the key".
+
+Downsides with this:
+1. A log can be instructed to look up arbitrary TXT records
+2. No versioning
+
+As far as we know there are no amplification threats with (1), but ideally it
+would only be possible to query TXT RRs that are actually relevant for Sigsum.
+
+Not having any versioning could potentially become a headache. All other log
+endpoints are versioned. There is no good reason to not have versioning here,
+unless that would imply something like registering many different things with
+IANA as a result.
+
+Proposal
+---
+Require that a domain hint is formatted as:
+
+ _sigsum_v0.*
+
+Examples of valid domain hints:
+
+ _sigsum_v0.com
+ _sigsum_v0.example.com
+ _sigsum_v0.sub.example.com
+
+Examples of invalid domain hints:
+
+ _sigsum_v0hello.example.com
+
+This change addresses both (1) and (2), without making DNS configs harder.
+
+Notes
+---
+For v1 we need to consider if something should be registered with IANA. Credit
+to Patrik Wallström who pointed us towards documentation about labels with
+underscores:
+ * https://www.rfc-editor.org/rfc/rfc8552.html
+ * https://www.iana.org/assignments/dns-parameters/dns-parameters.xhtml#underscored-globally-scoped-dns-node-names
+
+Note also that the dependency on TXT look-ups means that a "hidden log" via Tor
+would need help from a resolver that is also available over Tor (preferably an
+onion but at minimum reachable over TCP). This is because TXT records cannot be
+resolved over Tor. This proposal allows the used resolver to be restricted to
+only resolve _sigsum_*.
diff --git a/doc/proposals/2022-01-get-endpoints b/doc/proposals/2022-01-get-endpoints
new file mode 100644
index 0000000..cbe3170
--- /dev/null
+++ b/doc/proposals/2022-01-get-endpoints
@@ -0,0 +1,46 @@
+Proposal: change get-* endpoints that use HTTP post
+
+Background
+---
+Right now we HTTP POST ASCII key-value pairs on these endpoints:
+ * get-leaves
+ * get-inclusion-proof
+ * get-consistency-proof
+
+The original reason was to not have an additional parser, say, input-parameters
+with percent-encoding as part of the request URL.
+
+A major problem with this approach is that it will not be possible to benefit
+from HTTP caching. Debugging, with "URLs that reference data" also becomes more
+messy. You would have to say "I did printf <stuff> | curl ...".
+
+Proposal
+---
+Change these endpoints so that they use HTTP GET. Encode input params in URL:
+
+ <url>/get-leaves/10/20 # get leaves 10,11,...20
+ <url>/get-consistency-proof/10/20 # proof from tree size 10 to 20
+ <url>/get-inclusion-proof/10/<leaf hash in hex> # proof for tree size 10
+
+This notably avoids percent-encoding which is more messy.
+
+Notes
+---
+We considered if it would be a good idea to re-use our ASCII parser for the
+portion of the URL that encodes input data. The basic idea would be that
+different "end of key" and "end of value" patterns could be used that are better
+suited for a URL.
+
+For example, instead of (=,\n) one could use ([,]) as ("end of key", "end of value").
+ * get-leaves/start_size[10]end_size[12]
+ * get-consistency-proof/old_size[12]new_size[14]
+ * get-inclusion-proof/tree_size[10]leaf_hash[ab...ef]
+
+The reasons why we aborted this direction:
+ * We can not think of any concrete security risk with the shorter '/' proposal.
+ * There are very few parameters at play here, hard to confuse and quick
+ feedback loop if you do. For example, "Error=start size must be smaller or
+ equal to end size".
+ * We can be sure that the '/' proposal will not introduce any wonky
+ interoperability issues; picking a ("end of key", "end of value") would
+ require much more care.
diff --git a/doc/proposals/2022-01-log-url b/doc/proposals/2022-01-log-url
new file mode 100644
index 0000000..598fb43
--- /dev/null
+++ b/doc/proposals/2022-01-log-url
@@ -0,0 +1,26 @@
+Redefine "base URL" as "log URL" proposal
+
+Background
+---
+The current api.md specification requires that a log has a fixed unique "base
+URL". It is any valid HTTP(S) URL that can end with "/sigsum/v0/<endpoint>".
+
+Proposal
+---
+Remove the term "base URL" and instead define "log URL". A log URL is a valid
+HTTP(S) URL that ends with "/sigsum/v0/". Example of a valid log URL:
+
+ https://example.com:4711/opposum/sigsum/v0/
+
+This means that a named sigsum endpoint can be appended to a log's URL. For
+example, if the endpoint is "get-tree-head-quickly" the resulting "endpoint URL"
+would be:
+
+ https://example.com:4711/opposum/sigsum/v0/get-tree-head-quickly/
+
+And with input parameters for "get-leaves":
+
+ https://example.com:4711/opposum/sigsum/v0/get-leaves/42/4711/
+
+Note the final slash in all of the above URLs. Should that be enforced (?).
+ * XXX: Need to check in URL specification(s). Defer for now.
diff --git a/doc/proposals/2022-01-tree-head-endpoints b/doc/proposals/2022-01-tree-head-endpoints
new file mode 100644
index 0000000..b2831bf
--- /dev/null
+++ b/doc/proposals/2022-01-tree-head-endpoints
@@ -0,0 +1,118 @@
+Proposal: change tree-head endpoints
+
+Background
+---
+Right now the get-tree-head-to-sign endpoint returns the signed tree head that
+witnesses should cosign. It does not return any cosignatures. One needs to
+wait until the to-sign tree head is finalized and served via
+get-tree-head-cosigned. We also have a get-tree-head-latest endpoint that is
+sort of hanging around for "debug purposes".
+
+It would be nice if a submitter could find required cosignatures without always
+having to wait for five minutes. The log will likely have received a majority
+of cosignatures after one minute, but a submitter currently needs to wait the
+full duration before getting access via the get-tree-head-cosigned endpoint.
+
+It would also be nice to consider if the get-tree-head-latest endpoint can be
+removed.
+
+Here is a rough break-down of how we think about the sigsum API's usage via
+roles:
+ * Submitter
+ * add-leaf, until HTTP status 200 OK which should mean "you have been sequenced".
+ * [fetching an inclusion proof for a signed tree head to "verify sequencing"
+ is not a recommended usage pattern, and does not prevent DoS. The only
+ difference is that the submitter would notice that the log has not
+ included with regards to the latest tree head sooner than with regards
+ to the cosigned tree head. In both cases, there is no proof that
+ submitter got 200 OK without getting sequenced.]
+ * Distributor
+ * get-tree-head-cosigned
+ * get-inclusion-proof
+ * [wants "enough" cosignatures, sooner rather than later is a soft requirement]
+ * Monitor
+ * get-leaves
+ * get-tree-head-cosigned
+ * might hit get-{consistency,inclusion}-proof depending on implementation
+ * [wants as many cosignatures as possible, does not care about ~minutes of waiting]
+ * Witness
+ * get-consistency-proof
+ * get-tree-head-to-sign
+ * add-cosignature
+ * [does not / should not care about other cosignatures; just that the
+ log signed and that the tree head is consistent with prior history as
+ observed by the witness]
+ * End-user
+ * [does not hit any of the log's endpoints]
+ * "The curious"
+ * the latest signed tree head, as fast as possible for quick debug
+ probably. "is the thing I'm doing working".
+ * the latest cosigned tree head, with as many cosignatures as possible
+ for archiving
+
+Keep in mind that the below proposal should not introduce the log's key hash as
+output on any API endpoint. We removed this and other redundant output because
+that reduces the risk of faulty implementations that operate on untrusted input.
+
+For example, in the same way that a faulty witness could verify "the wrong
+consistency proof" if it just verified the proof against the tree sizes that the
+log returned redundantly (as opposed to the tree sizes the witness asked for), a
+faulty witness could end-up cosigning a tree head with another log's context
+because "they just copied the key hash and used it because it was there".
+
+Note that we cannot add the key_hash and cosignature fields to the output of
+get-tree-head-to-sign. Our ASCII parser is so simple that it does not permit
+empty lists. So, we will either need a way to handle empty lists, or tweak our
+endpoints so that they still do what we want without having any empty list.
+
+[Both rgdd and ln5 would like to avoid complicating the ASCII parser.]
+
+Proposal
+---
+1. Remove the get-tree-head-latest endpoint. We no longer have any recommended
+usage-pattern for this endpoint and so it should be removed. Our strongest
+arguments for removal are "don't use a signed tree head, it is sort of like a
+promise", and "it does not even help you prove that the log's HTTP status 200 OK
+semantics were faulty".
+2. The get-tree-head-to-sign endpoint is kept as is, but renamed.
+ * Purpose: used by witnesses.
+3. Add an endpoint that returns the logs "to-cosign" tree head and all
+cosignatures that were collected thus far. If no cosignatures were received
+yet, return an error to avoid having an empty list as output.
+ * Purpose: used by distributors, but could also be used by a witness'
+ internal monitoring setup ("is my witness working, are the signatures really
+ showing up?").
+4. Keep an endpoint that serves the "finalized" cosigned tree head.
+ * Purpose: mainly used by monitors, but could also be used by distributor's
+ that don't mind the additional waiting or by parties that want to archive
+ cosigned tree heads.
+
+This proposal currently does not have a name for the above endpoints. Help
+wanted.
+
+Notes
+---
+A witness polls the "get-tree-head-to-sign" endpoint as before. Witnesses are
+recommended to poll the log at least once per minute at randomly selected times.
+
+After a successful add-cosignature request, a witness should not attempt to add
+the same cosignature again. A log can refresh their "to-sign tree head" to
+instruct witnesses to send their cosignatures again for the same tree size.
+
+A witness operator may check that their cosignatures appear on the
+"get-tree-head-cosigned endpoints". Such checking would likely be part of how
+the operator monitors that the witness operates correctly (i.e., it would not be
+something that the witness software does itself after a successful
+add-cosignature request).
+
+A submitter ("Signer" in Figure 1) that wants a cosigned tree head that
+satisifies a given policy as fast as possible can poll the "dynamic cosigned
+tree head endpoint". Keep in mind that polling more than a few times per minute
+would not let you obtain cosignatures much faster, see the above recommendation
+for how often witnesses should provide their cosignatures.
+
+A helpful reflection with regards to naming:
+ * "The log's to-sign STH shows up, it gets filled-up with cosignatures; the
+ previous cosigned tree head is served on a separate endpoint. Then
+ "prev=curr, curr=new". I.e., there is a time aspect here that might be
+ helpful for naming, although previous and current would be bad choices."