Ugh! We definitely cannot have a model where a 10K JSON document is exploded
into 2MB worth of KV data. I’ve tried several times to follow the math here but
I’m failing. I can’t even get past this first bit:
> - each document is around 10Kb
> - each document consists of 1K of unique JSON paths
> - each document has 100 unique JSON field names
> - every scalar value is 100 bytes
If each document has 1000 paths, and each path (which leads to a unique scalar
value, right?) has a value of 100 bytes associated with it … how is the
document 10KB? Wouldn’t it need to be at least 100KB just by adding up all the
scalar values?
Adam
> On Feb 4, 2019, at 6:08 AM, Ilya Khlopotov <iil...@apache.org> wrote:
>
> Hi Michael,
>
>> For example, hears a crazy thought:
>> Map every distinct occurence of a key/value instance through a crypto hash
>> function to get a set of hashes.
>>
>> These can be be precomputed by Couch without any lookups in FDB. These
>> will be spread all over kingdom come in FDB and not lend themselves to
>> range search well.
>>
>> So what you do is index them for frequency of occurring in the same set.
>> In essence, you 'bucket them' statistically, and that bucket id becomes a
>> key prefix. A crypto hash value can be copied into more than one bucket.
>> The {bucket_id}/{cryptohash} becomes a {val_id}
>
>> When writing a document, Couch submits the list/array of cryptohash values
>> it computed to FDB and gets back the corresponding {val_id} (the id with
>> the bucket prefixed). This can get somewhat expensive if there's always a
>> lot of app local cache misses.
>>
>> A document's value is then a series of {val_id} arrays up to 100k per
>> segment.
>>
>> When retrieving a document, you get the val_ids, find the distinct buckets
>> and min/max entries for this doc, and then parallel query each bucket while
>> reconstructing the document.
>
> Interesting idea. Let's try to think it through to see if we can make it
> viable.
> Let's go through hypothetical example. Input data for the example:
> - 1M of documents
> - each document is around 10Kb
> - each document consists of 1K of unique JSON paths
> - each document has 100 unique JSON field names
> - every scalar value is 100 bytes
> - 10% of unique JSON paths for every document already stored in database
> under different doc or different revision of the current one
> - we assume 3 independent copies for every key-value pair in FDB
> - our hash key size is 32 bytes
> - let's assume we can determine if key is already on the storage without
> doing query
> - 1% of paths is in cache (unrealistic value, in real live the percentage is
> lower)
> - every JSON field name is 20 bytes
> - every JSON path is 10 levels deep
> - document key prefix length is 50
> - every document has 10 revisions
> Let's estimate the storage requirements and size of data we need to transmit.
> The calculations are not exact.
> 1. storage_size_per_document (we cannot estimate exact numbers since we don't
> know how FDB stores it)
> - 10 * ((10Kb - (10Kb * 10%)) + (1K - (1K * 10%)) * 32 bytes) = 38Kb * 10 *
> 3 = 1140 Kb (11x)
> 2. number of independent keys to retrieve on document read (non-range
> queries) per document
> - 1K - (1K * 1%) = 990
> 3. number of range queries: 0
> 4. data to transmit on read: (1K - (1K * 1%)) * (100 bytes + 32 bytes) = 102
> Kb (10x)
> 5. read latency (we use 2ms per read based on numbers from
> https://apple.github.io/foundationdb/performance.html)
> - sequential: 990*2ms = 1980ms
> - range: 0
> Let's compare these numbers with initial proposal (flattened JSON docs
> without global schema and without cache)
> 1. storage_size_per_document
> - mapping table size: 100 * (20 + 4(integer size)) = 2400 bytes
> - key size: (10 * (4 + 1(delimiter))) + 50 = 100 bytes
> - storage_size_per_document: 2.4K*10 + 100*1K*10 + 1K*100*10 = 2024K = 1976
> Kb * 3 = 5930 Kb (59.3x)
> 2. number of independent keys to retrieve: 0-2 (depending on index structure)
> 3. number of range queries: 1 (1001 of keys in result)
> 4. data to transmit on read: 24K + 1000*100 + 1000*100 = 23.6 Kb (2.4x)
> 5. read latency (we use 2ms per read based on numbers from
> https://apple.github.io/foundationdb/performance.html and estimate range read
> performance based on numbers from
> https://apple.github.io/foundationdb/benchmarking.html#single-core-read-test)
> - range read performance: Given read performance is about 305,000
> reads/second and range performance 3,600,000 keys/second we estimate range
> performance to be 11.8x compared to read performance. If read performance is
> 2ms than range performance is 0.169ms (which is hard to believe).
> - sequential: 2 * 2 = 4ms
> - range: 0.169
>
> It looks like we are dealing with a tradeoff:
> - Map every distinct occurrence of a key/value instance through a crypto hash:
> - 5.39x more disk space efficient
> - 474x slower
> - flattened JSON model
> - 5.39x less efficient in disk space
> - 474x faster
>
> In any case this unscientific exercise was very helpful. Since it uncovered
> the high cost in terms of disk space. 59.3x of original disk size is too much
> IMO.
>
> Are the any ways we can make Michael's model more performant?
>
> Also I don't quite understand few aspects of the global hash table proposal:
>
> 1. > - Map every distinct occurence of a key/value instance through a crypto
> hash function to get a set of hashes.
> I think we are talking only about scalar values here? I.e. `"#/foo.bar.baz":
> 123`
> Since I don't know how we can make it work for all possible JSON paths
> `{"foo": {"bar": {"size": 12, "baz": 123}}}":
> - foo
> - foo.bar
> - foo.bar.baz
>
> 2. how to delete documents
>
> Best regards,
> ILYA
>
>
> On 2019/01/30 23:33:22, Michael Fair <mich...@daclubhouse.net> wrote:
>> On Wed, Jan 30, 2019, 12:57 PM Adam Kocoloski <kocol...@apache.org wrote:
>>
>>> Hi Michael,
>>>
>>>> The trivial fix is to use DOCID/REVISIONID as DOC_KEY.
>>>
>>> Yes that’s definitely one way to address storage of edit conflicts. I
>>> think there are other, more compact representations that we can explore if
>>> we have this “exploded” data model where each scalar value maps to an
>>> individual KV pair.
>>
>>
>> I agree, as I mentioned on the original thread, I see a scheme, that
>> handles both conflicts and revisions, where you only have to store the most
>> recent change to a field. Like you suggested, multiple revisions can share
>> a key. Which in my mind's eye further begs the conflicts/revisions
>> discussion along with the working within the limits discussion because it
>> seems to me they are all intrinsically related as a "feature".
>>
>> Saying 'We'll break documents up into roughly 80k segments', then trying to
>> overlay some kind of field sharing scheme for revisions/conflicts doesn't
>> seem like it will work.
>>
>> I probably should have left out the trivial fix proposal as I don't think
>> it's a feasible solution to actually use.
>>
>> The comment is more regarding that I do not see how this thread can escape
>> including how to store/retrieve conflicts/revisions.
>>
>> For instance, the 'doc as individual fields' proposal lends itself to value
>> sharing across mutiple documents (and I don't just mean revisions of the
>> same doc, I mean the same key/value instance could be shared for every
>> document).
>> However that's not really relevant if we're not considering the amount of
>> shared information across documents in the storage scheme.
>>
>> Simply storing documents in <100k segments (perhaps in some kind of
>> compressed binary representation) to deal with that FDB limit seems fine.
>> The only reason to consider doing something else is because of its impact
>> to indexing, searches, reduce functions, revisions, on-disk size impact,
>> etc.
>>
>>
>>
>>>> I'm assuming the process will flatten the key paths of the document into
>>> an array and then request the value of each key as multiple parallel
>>> queries against FDB at once
>>>
>>> Ah, I think this is not one of Ilya’s assumptions. He’s trying to design a
>>> model which allows the retrieval of a document with a single range read,
>>> which is a good goal in my opinion.
>>>
>>
>> I am not sure I agree.
>>
>> Think of bitTorrent, a single range read should pull back the structure of
>> the document (the pieces to fetch), but not necessarily the whole document.
>>
>> What if you already have a bunch of pieces in common with other documents
>> locally (a repeated header/footer/ or type for example); and you only need
>> to get a few pieces of data you don't already have?
>>
>> The real goal to Couch I see is to treat your document set like the
>> collection of structured information that it is. In some respects like an
>> extension of your application's heap space for structured objects and
>> efficiently querying that collection to get back subsets of the data.
>>
>> Otherwise it seems more like a slightly upgraded file system plus a fancy
>> grep/find like feature...
>>
>> The best way I see to unlock more features/power is to a move towards a
>> more granular and efficient way to store and retrieve the scalar values...
>>
>>
>>
>> For example, hears a crazy thought:
>> Map every distinct occurence of a key/value instance through a crypto hash
>> function to get a set of hashes.
>>
>> These can be be precomputed by Couch without any lookups in FDB. These
>> will be spread all over kingdom come in FDB and not lend themselves to
>> range search well.
>>
>> So what you do is index them for frequency of occurring in the same set.
>> In essence, you 'bucket them' statistically, and that bucket id becomes a
>> key prefix. A crypto hash value can be copied into more than one bucket.
>> The {bucket_id}/{cryptohash} becomes a {val_id}
>>
>> When writing a document, Couch submits the list/array of cryptohash values
>> it computed to FDB and gets back the corresponding {val_id} (the id with
>> the bucket prefixed). This can get somewhat expensive if there's always a
>> lot of app local cache misses.
>>
>>
>> A document's value is then a series of {val_id} arrays up to 100k per
>> segment.
>>
>> When retrieving a document, you get the val_ids, find the distinct buckets
>> and min/max entries for this doc, and then parallel query each bucket while
>> reconstructing the document.
>>
>> The values returned from the buckets query are the key/value strings
>> required to reassemble this document.
>>
>>
>> ----------
>> I put this forward primarily to hilite the idea that trying to match the
>> storage representation of documents in a straight forward way to FDB keys
>> to reduce query count might not be the most performance oriented approach.
>>
>> I'd much prefer a storage approach that reduced data duplication and
>> enabled fast sub-document queries.
>>
>>
>> This clearly falls in the realm of what people want the 'use case' of Couch
>> to be/become. By giving Couch more access to sub-document queries, I could
>> eventually see queries as complicated as GraphQL submitted to Couch and
>> pulling back ad-hoc aggregated data across multiple documents in a single
>> application layer request.
>>
>> Hehe - one way to look at the database of Couch documents is that they are
>> all conflict revisions of the single root empty document. What I mean be
>> this is consider thinking of the entire document store as one giant DAG of
>> key/value pairs. How even separate documents are still typically related to
>> each other. For most applications there is a tremendous amount of data
>> redundancy between docs and especially between revisions of those docs...
>>
>>
>>
>> And all this is a long way of saying "I think there could be a lot of value
>> in assuming documents are 'assembled' from multiple queries to FDB, with
>> local caching, instead of simply retrieved"
>>
>> Thanks, I hope I'm not the only outlier here thinking this way!?
>>
>> Mike :-)
>>
