Difference between revisions of "Development Team/Datastore Rewrite"

From Sugar Labs
Jump to navigation Jump to search
 
(22 intermediate revisions by 6 users not shown)
Line 1: Line 1:
 +
<noinclude>{{ GoogleTrans-en | es =show | bg =show | zh-CN =show | zh-TW =show | hr =show | cs =show | da =show | nl =show | fi =show | fr =show | de =show | el =show | hi =show | it =show | ja =show | ko =show | no =show | pl =show | pt =show | ro =show | ru =show | sv =show }}</noinclude>
 +
{{TOCright}}
 +
 +
== Introduction ==
 +
 +
This page describes the design of a new DataStore implementation that was shipped first in the [[0.84]] (April 2009) Sugar release.
 +
 
== Goals ==
 
== Goals ==
  
Line 35: Line 42:
 
By examining the directory structure, we know where is localized the data related to each entry. We don't depend any more on a binary structure that could become corrupted and unusable as a whole.
 
By examining the directory structure, we know where is localized the data related to each entry. We don't depend any more on a binary structure that could become corrupted and unusable as a whole.
  
=== Metadata is stored as a single file for each entry ===
+
=== Metadata is stored in a single file per property ===
  
Each entry has its metadata stored in a single file, so that if corruption happened on one of those, the rest of the entries would be unaffected. The format in which metadata is stored even allows to recover from a malformed property by just dropping it.
+
Metadata for each entry is stored in several files, one per property. In this way, if corruption happened on one those properties, the rest of the entry (and the other entries in the DS) would be unaffected.
 +
: See [[olpc:Low-level_Activity_API#Meta_Data]].
  
 
=== Queries are accelerated with a disposable database ===
 
=== Queries are accelerated with a disposable database ===
Line 46: Line 54:
  
 
This improves storage efficiency in general, but in our case is more important because we wish to record in the journal several interactions that refer to the same file. For example, "Downloaded lesson3.pdf", "Read lesson3.pdf", "Sent lesson3.pdf to Juan" would all refer to the same file and we need to only store it once.
 
This improves storage efficiency in general, but in our case is more important because we wish to record in the journal several interactions that refer to the same file. For example, "Downloaded lesson3.pdf", "Read lesson3.pdf", "Sent lesson3.pdf to Juan" would all refer to the same file and we need to only store it once.
 +
 +
== Layout on disk ==
 +
 +
The proposed implementation relies heavily on the data structures provided by the filesystem, thus the layout in which files are disposed on disk is a fundamental part of its design.
 +
 +
Example of a datastore containing 5 entries, two of them referring to the same file (with checksum 464493d8d929436b6152e868867ed451):
 +
 +
1a
 +
      1ab88287-766a-4d98-a7c0-4233dc48647a
 +
            data
 +
            metadata
 +
                  uid
 +
                  checksum
 +
                  activity_id
 +
                  mime_type
 +
                  preview
 +
                  share-scope
 +
                  timestamp
 +
                  title
 +
2b
 +
      2b90597c-0912-4e7f-8eeb-71a0f004490d
 +
            data
 +
            metadata
 +
                  uid
 +
                  checksum
 +
                  activity_id
 +
                  mime_type
 +
                  preview
 +
                  share-scope
 +
                  timestamp
 +
                  title
 +
3c
 +
      3cdf5f0e-7595-4166-b1f9-cbedfcfe1c4a
 +
            data -> 2b/2b90597c-0912-4e7f-8eeb-71a0f004490d/data
 +
            metadata
 +
                  uid
 +
                  checksum
 +
                  activity_id
 +
                  mime_type
 +
                  preview
 +
                  share-scope
 +
                  timestamp
 +
                  title
 +
4d
 +
      4db11d29-2f07-4452-bd8e-22a6a483ac19
 +
            data
 +
            metadata
 +
                  uid
 +
                  checksum
 +
                  activity_id
 +
                  mime_type
 +
                  preview
 +
                  share-scope
 +
                  timestamp
 +
                  title
 +
      4d9f2027-b41e-4015-a848-6b3972193eb8
 +
            data
 +
            metadata
 +
                  uid
 +
                  checksum
 +
                  activity_id
 +
                  mime_type
 +
                  preview
 +
                  share-scope
 +
                  timestamp
 +
                  title
 +
checksums
 +
      464493d8d929436b6152e868867ed451
 +
            2b90597c-0912-4e7f-8eeb-71a0f004490d
 +
            3cdf5f0e-7595-4166-b1f9-cbedfcfe1c4a
 +
index
 +
      flintlock
 +
      iamflint
 +
      postlist.baseA
 +
      postlist.baseB
 +
      postlist.DB
 +
      record.baseA
 +
      record.baseB
 +
      record.DB
 +
      termlist.baseA
 +
      termlist.baseB
 +
      termlist.DB
 +
      value.baseA
 +
      value.baseB
 +
      value.DB
 +
index_updated
 +
version
 +
 +
'''1a''': directory holding entries, it's only function is to avoid having too many directories in a single directory, as this is considered specially harmful on jffs2.
 +
 +
'''1a/1ab88287-...-4233dc48647a''': directory holding the files related to one entry
 +
 +
'''1a/1ab88287-...-4233dc48647a/data''': file related to an entry
 +
 +
'''1a/1ab88287-...-4233dc48647a/metadata''': directory containing a file for each metadata property of an entry
 +
 +
'''2b/2b90597c-...-71a0f004490d/metadata/activity_id''': file containing the value of the '''activity_id''' property
 +
 +
'''3c/3cdf5f0e-...-cbedfcfe1c4a/data''': hard link to the same file in the entry '''2b90597c-...-71a0f004490d'''
 +
 +
'''checksums''': directory containing a directory per each file contained in the DS, named by its md5 checksum
 +
 +
'''checksums/464493d8d929436b6152e868867ed451''': directory containing files named by each entry that contain a file with this checksum
 +
 +
'''checksums/464493d8d929436b6152e868867ed451/2b90597c-...-71a0f004490d''' file named by the uid of an entry the file of which has this checksum.
 +
 +
'''index''': directory containing all files that belong to the search database. Can be deleted and recreated from the rest of the DS if needed without incurring in data loss.
 +
 +
'''index_updated''': When this file is not present, is because the xapian index is being rebuilt because it cannot be opened. Queries should fall back to display all the files on disk until the index is rebuilt.
 +
 +
'''version''': file that contains the version of the file layout. Currently we are at 1. Is updated when an earlier version is migrated to a newer one.
 +
 +
== Source code ==
 +
 +
http://git.sugarlabs.org/projects/sugar-datastore

Latest revision as of 21:33, 23 February 2010

Introduction

This page describes the design of a new DataStore implementation that was shipped first in the 0.84 (April 2009) Sugar release.

Goals

Reliability

A good DataStore doesn't lose data easily.

Performance

Queries should be fast enough for the journal to be very responsive when browsing its contents.

Activities should be able to store their data quickly and present a fast UI to their users.

The shell should be able to quickly query the DS to allow the user to resume entries from other views than the journal.

Maintainability

The original implementation tried to achieve goals that were hard and that proved not to be necessary at this stage. This has caused the code base to be unnecessarily complex and several changes to the requirements added considerable confusion to it. We wish to focus the code on what is really needed and do it well.

Custom metadata properties

Activities should be able to store in their entries the metadata they wish, should not be limited to a predefined set.

More efficient file storage

Identical files should be stored just once.

Versioned entries (not fulfilled yet)

Entries may be related in version trees.

Design

Filesystem knows which entries are stored

By examining the directory structure, we know where is localized the data related to each entry. We don't depend any more on a binary structure that could become corrupted and unusable as a whole.

Metadata is stored in a single file per property

Metadata for each entry is stored in several files, one per property. In this way, if corruption happened on one those properties, the rest of the entry (and the other entries in the DS) would be unaffected.

See olpc:Low-level_Activity_API#Meta_Data.

Queries are accelerated with a disposable database

This allows us to efficiently query the stored entries, but as we only use the database to accelerate queries, we can drop and recreate it in case of corruption or update to an incompatible database format.

Detect identical files and hard-link them

This improves storage efficiency in general, but in our case is more important because we wish to record in the journal several interactions that refer to the same file. For example, "Downloaded lesson3.pdf", "Read lesson3.pdf", "Sent lesson3.pdf to Juan" would all refer to the same file and we need to only store it once.

Layout on disk

The proposed implementation relies heavily on the data structures provided by the filesystem, thus the layout in which files are disposed on disk is a fundamental part of its design.

Example of a datastore containing 5 entries, two of them referring to the same file (with checksum 464493d8d929436b6152e868867ed451):

1a
      1ab88287-766a-4d98-a7c0-4233dc48647a
            data
            metadata
                  uid
                  checksum
                  activity_id
                  mime_type
                  preview
                  share-scope
                  timestamp
                  title
2b
      2b90597c-0912-4e7f-8eeb-71a0f004490d
            data
            metadata
                  uid
                  checksum
                  activity_id
                  mime_type
                  preview
                  share-scope
                  timestamp
                  title
3c
      3cdf5f0e-7595-4166-b1f9-cbedfcfe1c4a
            data -> 2b/2b90597c-0912-4e7f-8eeb-71a0f004490d/data
            metadata
                  uid
                  checksum
                  activity_id
                  mime_type
                  preview
                  share-scope
                  timestamp
                  title
4d
      4db11d29-2f07-4452-bd8e-22a6a483ac19
            data
            metadata
                  uid
                  checksum
                  activity_id
                  mime_type
                  preview
                  share-scope
                  timestamp
                  title
      4d9f2027-b41e-4015-a848-6b3972193eb8
            data
            metadata
                  uid
                  checksum
                  activity_id
                  mime_type
                  preview
                  share-scope
                  timestamp
                  title
checksums
      464493d8d929436b6152e868867ed451
            2b90597c-0912-4e7f-8eeb-71a0f004490d
            3cdf5f0e-7595-4166-b1f9-cbedfcfe1c4a
index
      flintlock
      iamflint
      postlist.baseA
      postlist.baseB
      postlist.DB
      record.baseA
      record.baseB
      record.DB
      termlist.baseA
      termlist.baseB
      termlist.DB
      value.baseA
      value.baseB
      value.DB
index_updated
version

1a: directory holding entries, it's only function is to avoid having too many directories in a single directory, as this is considered specially harmful on jffs2.

1a/1ab88287-...-4233dc48647a: directory holding the files related to one entry

1a/1ab88287-...-4233dc48647a/data: file related to an entry

1a/1ab88287-...-4233dc48647a/metadata: directory containing a file for each metadata property of an entry

2b/2b90597c-...-71a0f004490d/metadata/activity_id: file containing the value of the activity_id property

3c/3cdf5f0e-...-cbedfcfe1c4a/data: hard link to the same file in the entry 2b90597c-...-71a0f004490d

checksums: directory containing a directory per each file contained in the DS, named by its md5 checksum

checksums/464493d8d929436b6152e868867ed451: directory containing files named by each entry that contain a file with this checksum

checksums/464493d8d929436b6152e868867ed451/2b90597c-...-71a0f004490d file named by the uid of an entry the file of which has this checksum.

index: directory containing all files that belong to the search database. Can be deleted and recreated from the rest of the DS if needed without incurring in data loss.

index_updated: When this file is not present, is because the xapian index is being rebuilt because it cannot be opened. Queries should fall back to display all the files on disk until the index is rebuilt.

version: file that contains the version of the file layout. Currently we are at 1. Is updated when an earlier version is migrated to a newer one.

Source code

http://git.sugarlabs.org/projects/sugar-datastore