Hi Iceberg Community,
This is a great conversation so far, and thanks everyone for the valuable
inputs!
I'd like to articulate 2 things that we have to keep in mind with the
design:
*1: There are 2 interfaces here that we should consider:*
What I mean by this is that so far we have been talking about the REST
spec, more narrowly the HTTP communication between Iceberg's REST client
and the REST server. I think the proposed solution with the ETag absolutely
makes sense within this context.
However, the usual way of a client interacting with an Iceberg catalog
(including REST) is the Catalog API in the library. This API offers a
loadTable(TableIdentifier) function that returns a Table object. With the
above HTTP-based solution in mind I don't think we could give any
meaningful results if the HTTP layer finds that the table hasn't changed. I
argued already against pushing the caching responsibilities from the
clients into the HTTP layer (mostly because of losing the control over the
cache, and also observability won't be straightforward) so let's assume for
now that we won't do caching in the HTTP layer, only execute the loadTable
calls to the REST catalog by setting the ETag. In case we get a 304 we
won't be able to construct a Table object to answer the
Catalog.loadTable(TableIdentifier) call. We could return null or throw an
exception but I don't find any of them appropriate.
*2: There are catalog types other than REST*
I started this conversation focusing on the REST spec, but the more I think
of this the more I feel that the same functionality should be offered for
all the other catalog types too. Let's assume that we have an engine that
caches table metadata and initially uses REST catalog. For such an engine
the proposed solution would solve the problem of checking table freshnes
and also reloading the table metadata. A simple code for that could be
enough if we configured our HTTP client properly (just sketched a simple
example):
tableCache_.put(catalog_.loadTable(tableIdentifier));
Also let's assume we solve the issue in 1) and we can answer such a call
even if we get 304 from the server as the table is unchanged. So with this
solution with the REST catalog we can be sure that the table is only loaded
from the catalog if changed (or the age expired). But what if we configure
another catalog, let's say HiveCatalog. The very same code for that catalog
would trigger a table reload for every execution causing unexpected
performance issues.
I have to double check but I assume that this HTTP approach wouldn't be
feasible for other catalog types unfortunately.
I hope these arguments make sense :)
*As a partial solution this is what I have in mind:*
We can add another function into the catalog API for this purpose. Let's
say something like this:
Table loadTable(Table existingTable);
What advantages I see with this:
- This could solve issue 1) above. In case the table hasn't changed we can
simply return 'existingTable' without using HTTP Cache.
- The clients wouldn't need to explicitly call for isLatest() and such
functions to check for freshness, and they wouldn't need to trigger table
reloading for themselve. This API would be expected to cover this under the
hood.
- The current Catalog.loadTable(TableIdentifier) API wouldn't be enough for
all the catalog types on it's own, but with this one each catalog
implementations (e.g. HiveCatalog, REST catalog, etc.) then can implement
their own way of doing freshness checks and table reloads. For REST we
could follow the HTTP ETag approach, while for other catalogs we could
follow other approaches.
Regards,
Gabor
On Mon, Nov 18, 2024 at 8:48 AM Shani Elharrar <sh...@upsolver.com.invalid>
wrote:
> You're totally right. Perhaps using a "Content-Location" header might be a
> better fit for that.
>
> Shani.
>
> On 18 Nov 2024, at 9:27, Taeyun Kim <taeyun....@innowireless.com> wrote:
>
>
> Hi,
>
> Here are my thoughts:
>
> - The value of ETag is (as far as I know) defined as an opaque string by
> the specification, meaning the client shouldn’t interpret or assign any
> significance to it, regardless of what the server specifies. It’s best to
> avoid the client giving any particular meaning to the ETag value.
> - One major advantage of the header approach compared to other methods is
> that if an update has occurred, the updated content can be immediately
> included in the response without requiring an additional request. This
> saves one request-response round-trip (although It’s also possible to
> define a separate endpoint with the same functionality).
> - Since the Iceberg REST catalog server is effectively a type of HTTP
> server, at least in theory, it may be expected to handle HTTP cache and
> validation-related processes. The header approach can be seen as leveraging
> this mechanism appropriately.
> - The header approach doesn’t have to be limited to the
> /v1/{prefix}/namespaces/{namespace}/tables/{table} endpoint. It could also
> be applied to all GET-based endpoints, though this might broaden the scope
> significantly.
>
> Thank you.
>
>
>
> -----Original Message-----
> From: "Shani Elharrar" <sh...@upsolver.com.invalid>
> To: <dev@iceberg.apache.org>;
> Cc: <dev@iceberg.apache.org>;
> Sent: 2024-11-18 (월) 16:21:16 (UTC+09:00)
> Subject: Re: [DISCUSS] REST: Way to query if metadata pointer is the latest
>
> Using the metadata file name as ETag is nice way to go. In that case,
> adding HEAD method support to the loadTable endpoint will return the latest
> metadata pointer, which can be used to support "isLatest" without returning
> the body. It can be also leveraged in order to return the latest metadata
> location of the table.
>
> Shani.
>
> On 18 Nov 2024, at 8:52, Yufei Gu <flyrain...@gmail.com> wrote:
>
>
>
> Hi Taeyun,
>
> Thank you for the clear explanation.
>
> I agree that the ETag solution is more suitable. If we were going that
> way, I'd propose a customized version number as an ETag—for instance,
> leveraging the metadata.json file name as the identifier.
>
> To summarize, HTTP caching relies on headers (e.g., ETag or Last-Modified)
> to validate whether a version is up-to-date, whereas the alternative
> approach proposed above uses an additional parameter for verification. From
> my perspective, there isn’t a fundamental difference between the two, so
> I’m OK with either.
>
> A couple of points to note:
>
> 1. Both approaches would require changes to the "loadTable" endpoint.
> 2. A minor advantage of HTTP caching is that it integrates seamlessly
> with browsers, but since most clients of the Iceberg REST catalog aren’t
> browsers, this may not be a significant factor.
> 3. I’d also recommend considering the requirement to retrieve multiple
> tables(e.g., all tables under a namespace, or a list of table names) from
> the catalog. This requires a new endpoint and may not work with HTTP
> caching.
>
> Let me know your thoughts or if there’s anything else to consider.
> Yufei
>
>
> On Sun, Nov 17, 2024 at 6:43 PM Taeyun Kim <taeyun....@innowireless.com>
> wrote:
>
> Hi,
>
> To Gabor:
> It doesn’t seem necessary to interpret HTTP caching literally in this
> context.
> Simply using the HTTP headers defined by HTTP caching to check the
> freshness of metadata should be sufficient.
> There’s no requirement for the client to duplicate or store cached HTTP
> responses.
>
> To Yufei:
> As I understand it, the client doesn’t send its own timestamp but instead
> uses the timestamp originally provided by the server in the Last-Modified
> header.
> Therefore, clock synchronization issues should not be a concern.
>
> Here’s the general flow of HTTP cache validation based on
> If-Modified-Since:
>
> - Client: initial request:
>
> GET (url) HTTP/1.1
>
> - Server response:
>
> HTTP/1.1 200 OK
> Last-Modified: (date1)
> Cache-Control: no-store, no-cache, max-age=0, must-revalidate,
> proxy-revalidate
> (with response body)
>
> - Client: validation request:
>
> GET (url) HTTP/1.1
> If-Modified-Since: (date1)
>
> - Server response (if unchanged):
>
> HTTP/1.1 304 Not Modified
> Last-Modified: (date1)
> Cache-Control: no-store, no-cache, max-age=0, must-revalidate,
> proxy-revalidate
> (without response body)
>
> - Server response (if updated):
>
> HTTP/1.1 200 OK
> Last-Modified: (date2)
> Cache-Control: no-store, no-cache, max-age=0, must-revalidate,
> proxy-revalidate
> (with response body)
>
> However, using time-based freshness checks can present challenges, such as
> parsing time formats or synchronizing file update times across servers.
> To address these issues, HTTP cache validation based on ETag is also
> defined in the specification.
>
> Here’s the flow for ETag-based validation:
>
> - Client: initial request:
>
> GET (url) HTTP/1.1
>
> - Server response:
>
> HTTP/1.1 200 OK
> ETag: "(arbitrary string 1 generated by the server)"
> Cache-Control: no-store, no-cache, max-age=0, must-revalidate,
> proxy-revalidate
> (with response body)
>
> - Client: validation request:
>
> GET (url) HTTP/1.1
> If-None-Match: "(arbitrary string 1 generated by the server)"
>
> - Server response (if unchanged):
>
> HTTP/1.1 304 Not Modified
> ETag: "(arbitrary string 1 generated by the server)"
> Cache-Control: no-store, no-cache, max-age=0, must-revalidate,
> proxy-revalidate
> (without response body)
>
> - Server response (if updated):
>
> HTTP/1.1 200 OK
> ETag: "(arbitrary string 2 generated by the server)"
> Cache-Control: no-store, no-cache, max-age=0, must-revalidate,
> proxy-revalidate
> (with response body)
>
> The server can choose to use either If-Modified-Since or ETag for
> freshness validation.
> Alternatively, to simplify the implementation related to the Iceberg REST
> catalog, it might make sense to define only the more accurate ETag-based
> validation in the spec.
> For reference, RFC 9110 recommends specifying both ETag and Last-Modified.
> When both are provided, ETag takes precedence.
>
> Note on Cache-Control Headers:
> The Cache-Control values in the examples above are intended to ensure that
> the client validates freshness with the server on every request. Writing
> the header in this extended format is primarily to accommodate outdated
> HTTP/1.1 implementations. However, under the HTTP/1.1 specification, the
> following is sufficient:
>
> Cache-Control: no-cache
>
> That’s all for now.
> Thank you.
>
>
> -----Original Message-----
> From: "Yufei Gu" <flyrain...@gmail.com>
> To: <dev@iceberg.apache.org>;
> Cc:
> Sent: 2024-11-16 (토) 02:51:05 (UTC+09:00)
> Subject: Re: [DISCUSS] REST: Way to query if metadata pointer is the latest
>
>
>
> How does HTTP caching handle desynchronized clocks between clients and the
> server?
>
> At t0, the client gets the latest table version.
> At t1, the server makes a new commit.
> At t2, the client sends a request with a timestamp t2, but due to
> desynchronization, it refers to t0.
>
> The server may reply with 304 Not Modified, causing the client to think
> its cache is up-to-date and miss the commit at t1.
>
>
>
> Yufei
>
>
>
>
> On Fri, Nov 15, 2024 at 6:37 AM Gabor Kaszab <gaborkas...@apache.org>
> wrote:
> Hi All,
>
>
> First of all it's great to see that there are others who could benefit
> from giving a solution to this problem. I appreciate all the comments and
> feedback so far.
> There were a number of different opinions, so let me start with
> summarizing the different topics that came up:
>
>
> New endpoint vs using an existing endpoint:
> Based on the answers (Fokko, Yufei) I had the impression that we should be
> careful when adding new REST endpoints, and we should examine the re-use of
> existing endpoints first. Let's do that then, and in case we don't find it
> feasible then we can still fall back to any of my initial proposals
> (isLatest() or metadataLocation()).
>
>
> Granularity of freshness checks:
> It was brought up (Dmitri) that we might not want to do the metadata
> freshness checks solely based on metadata location, but we should consider
> doing more granular freshness checks. I personally don't see much benefit
> of designing this solution like that, TBH, but seeing some use-cases could
> help us understand the motivation here.
> Let me share my opinion on some of the arguments:
>
>
> "A change in metadata location does not necessarily mean a change in
> metadata content"
>
>
> AFAIK whenever Iceberg creates a new metadata file there is some change in
> the metadata itself. There might not be a new snapshot, though in the cases
> of e.g. a schema/partition evolution. But even in these cases triggering a
> table reload could make sense to me (e.g. answering SHOW CREATE TABLE and
> similar queries). Additionally, I'd assume the number of metadata location
> changes that don't create a new snapshot is too negligible to optimize for.
> Dmitri, let me know if I misunderstood something.
>
>
> "it may still be beneficial to permit the client to ask for changes to
> specific areas of metadata"
>
> This seems like a use-case that the partial metadata loading proposal
> could solve. To identify the need to load a specific part of the metadata
> with partial metadata loading seems an overkill to design with my proposal,
> if this is what you have in mind. Also I found that the partial metadata
> loading proposal faces serious headwinds, so I wouldn't rely on it at the
> moment.
>
>
> Re-using tableExists
> I think there is a consensus here that tableExists returning a metadata
> location could work but seems more like a workaround and could be
> misleading for the users.
>
> Partial metadata loading could solve this:
> (Yufei) I agree, it would be perfect for my use-case and I'm following the
> discussion on the proposal. However, for me it seems, as I wrote above,
> that the proposal faces serious headwinds now and I honestly wouldn't
> expect a solution in the short term. But solving the freshness problems is
> a more urgent thing to solve, not just for myself and Impala but apparently
> to many other stakeholders in the community according to the interest on
> this thread.
> Hence, I propose to come up with a separate solution for freshness checks,
> and we can still move to using partial metadata loading once that's out.
>
>
> Use HTTPCache and If-Modified-Since with loadTable
> This solution seems to do the trick for us. Let me do some research myself
> to see if there are any difficulties implementing this. Currently, I have
> more questions than answers wrt this approach :)
> - The initial problem is to answer freshness questions for the cached
> tables on the client side. If we introduce HttpCaching wouldn't we
> introduce the same problem but on a different level of representation. We'd
> then need to decide the freshness/staleness of the cached data in the HTTP
> layer.
> - If we cache the HTTP responses for a loadTable then we essentially cache
> the content of the metadata.jsons including the snapshot and metadata log
> and everything, plus the snapshot list (and I think the manifests for the
> latest snapshot). I believe that the size of this can easily reach the low
> megabytes range in memory, so in total keeping them in the HTTP Cache for
> all the tables we have queried can easily mean that we keep a couple of GBs
> in memory just for this purpose.
> For engines that already cache table metadata wouldn't this mean that we
> will cache some parts of the metadata redundantly?
> - How would we decide what is the max-age of a cached table metadata in
> the HTTP Cache? Would it be configurable so that each engine could use
> whatever it prefers?
>
>
> Sorry if any of the questions doesn't make sense, I just want to make sure
> I understand all the aspects of this approach.
>
>
> An additional topic I have in mind:
> REST catalog vs other catalogs:
> Now we are focusing our discussion on the REST spec, but I think it would
> be beneficial to extend our focus and cover other catalog implementations
> too. I don't think that this problem of data freshness is specific to REST
> catalog, it could affect any table in any other catalog too.
>
>
> I'll continue my investigation wrt the proposals, I just wanted to flush
> out and sum up what we have now before the weekend.
>
>
> Regards,
> Gabor
>
>
>
>
> On Fri, Nov 15, 2024 at 10:16 AM Jean-Baptiste Onofré <j...@nanthrax.net>
> wrote:
> Hi,
>
> I like the idea and it makes sense. As soon as it's clearly stated in
> the spec (using If-Modified-Since header and 304 status code), it
> looks good to me.
>
> Thanks !
> Regards
> JB
>
> On Fri, Nov 15, 2024 at 1:58 AM Taeyun Kim <taeyun....@innowireless.com>
> wrote:
> >
> > Hi,
> >
> > (Apologies if this email is a duplicate. This is my third attempt.)
> >
> > I also need a way to ensure that my table data is up-to-date. For now,
> I’m handling this by setting an expiration period after which I fetch the
> data again, regardless of its freshness.
> >
> > Here are my thoughts on the current suggestions. Please correct me if
> I've misunderstood any of the points.
> >
> > - isLatest(): This function could be inefficient since it would require
> an additional round-trip to fetch the metadata if it’s not up-to-date. This
> would result in two round-trips overall, which seems suboptimal.
> > - metadataLocation(): This has a similar issue as isLatest(). BTW,
> according to the REST catalog API documentation for LoadTableResult schema,
> it states, "Clients can check whether metadata has changed by comparing
> metadata locations after the table has been created." (
> https://github.com/apache/iceberg/blob/3659ded18d50206576985339bd55cd82f5e200cc/open-api/rest-catalog-open-api.yaml#L3175)
> This suggests that if the metadata location has changed, the metadata can
> be considered updated.
> > - tableExists(): Based on the name, this function seems to serve a
> different purpose.
> >
> > Here is my suggestion:
> >
> > Since HTTP has built-in caching features (
> https://developer.mozilla.org/en-US/docs/Web/HTTP/Caching), and REST
> catalogs operate over HTTP, it seems natural to leverage HTTP caching
> mechanisms. For example, HTTP includes the If-Modified-Since header and the
> 304 Not Modified status code. Using this approach, we could achieve data
> freshness with a single round-trip, fetching updated data only if there are
> modifications.
> >
> > What do you think about defining the spec in this direction?
> >
> > Thank you.
> >
> >
> >
> >
> > -----Original Message-----
> > From: "Yufei Gu" <flyrain...@gmail.com>
> > To: <dev@iceberg.apache.org>;
> > Cc:
> > Sent: 2024-11-13 (수) 03:43:24 (UTC+09:00)
> > Subject: Re: [DISCUSS] REST: Way to query if metadata pointer is the
> latest
> >
> >
> >
> > Hi Gamber,
> >
> > Thanks for the proposal! Impala isn’t unique in needing this—I've seen
> similar requirements from other engines.
> >
> > As others pointed out, using the “tableExists” endpoint seems like a
> workaround. I don't consider it a permanent way forward. We could address
> this by either modifying the current load table endpoint or introducing a
> new one, but ideally, we should avoid adding endpoints for every specific
> need. With that, partial metadata loading seems like a strong approach
> here, we will need certain agreement though. I'd suggest the community
> consider the use cases seriously. We need a way forward.
> >
> > I’m also not too concerned about using metadata file paths to verify the
> latest table version; clients can simply extract metadata filenames, which
> include the UUID.
> >
> > Yufei
> >
> >
> >
> >
> > On Tue, Nov 12, 2024 at 7:46 AM Jean-Baptiste Onofré <j...@nanthrax.net>
> wrote:
> >
> > Hi Fokko
> >
> > I like the idea, but I think it's more a workaround and could be
> > confusing for users :)
> >
> > Regards
> > JB
> >
> > On Tue, Nov 12, 2024 at 2:53 PM Fokko Driesprong <fo...@apache.org>
> wrote:
> > >
> > > Hey Gabor,
> > >
> > > Thanks for raising this. While reading this, my first thought is to
> leverage the `tableExists` operation:
> > >
> https://github.com/apache/iceberg/blob/e3f39972863f891481ad9f5a559ffef093976bd7/open-api/rest-catalog-open-api.yaml#L1129-L1160
> > >
> > > This doesn't return anything today, but we could return a payload to
> the latest metadata.json.
> > >
> > > Looking forward to what others think.
> > >
> > > Kind regards,
> > > Fokko
> > >
> > >
> > >
> > >
> > > Op di 12 nov 2024 om 14:33 schreef Shani Elharrar
> <sh...@upsolver.com.invalid>:
> > >>
> > >> I recommend option (b), provided there is no partial metadata
> loading. We implemented option (b) internally to facilitate partial
> metadata loading, as we have tables with hundreds of thousands of
> snapshots. This results in metadata that occupies approximately 500 MB in
> memory (excluding the JsonNodes), which is a significant load for some of
> our services.
> > >>
> > >> Shani.
> > >>
> > >> On 12 Nov 2024, at 14:12, Gabor Kaszab <gaborkas...@apache.org>
> wrote:
> > >>
> > >> Hey Iceberg Community,
> > >>
> > >> Background:
> > >> Impala is designed in a way to cache the Iceberg table metadata
> (BaseTable objects in practice) for faster access. Currently, Impala is
> tightly coupled with HMS and in turn with the HiveCatalog, and in order to
> keep the cached table objects up-to-date there is a notification mechanism
> driven by HMS to notify Impala about any changes in the table metadata.
> > >> The Impala community is actively looking for ways to decouple HMS
> from Impala and provide a way to use Impala without the need for HMS, and
> get the Iceberg table metadata from other catalog Implementations mainly
> focusing now on REST catalogs.
> > >>
> > >> Problem to solve:
> > >> We identified a particular missing functionality in the current REST
> spec: For engines that cache table metadata currently there is no way to
> check if that table metadata is up-to-date or not, and whether the engine
> should reload the metadata for that table or not without getting a whole
> table object from the catalog. For this I think the REST catalog (but in
> fact I think this could apply to any other catalogs) should be able to
> answer a question like:
> > >> "Hi Catalog, I have this version of this table, is it up-to-date?"
> > >>
> > >> Proposal:
> > >> I've been following the discussion about partial metadata loading
> that could be also used to answer the above question, but I have the
> impression now that the conversation stopped making any progress.
> > >> So instead of waiting for partial metadata loading I propose to have
> an addition to the REST spec now to answer the question I raised above:
> > >>
> > >> a) boolean isLatest(TableIdentifier ident, String metadataLocation);
> > >> b) String metadataLocation(TableIdentifier ident);
> > >>
> > >> Any of the above 2 approaches could help engines to decide if they
> have to invalidate/reload particular table metadata in the cache. I
> personally would go for option a) but would be open to hear other opinions.
> > >>
> > >> I'd like to know if the community could support me extending the REST
> spec with any of the 2 options.
> > >>
> > >> Regards,
> > >> Gabor
> > >>
> > >>
>
>
