Riak CS Release Notes

Riak S2 (Riak CS) 2.1.0 Release Notes

Released October 13, 2015.

This is a backwards-compatible* release that introduces a new metrics system, garbage collection refinements, and several other new features. Riak S2 2.1 is designed to work with both Riak KV 2.0.5+ and 2.1.1+.

Note: This release is backwards compatible only with the Riak S2 2.x series.

Riak KV 2.1.1 Usage Note

Riak KV 2.1.1 includes a copy of riak_cs_kv_multi_backend, therefore there is no need to add lines specifying special multi_backend and add_paths configurations in advanced.config.

Instead, you can set the following in riak.conf:

storage_backend = prefix_multi
cs_version = 20100

If you need storage calculation, you will still require the add_paths config to load MapReduce codes into Riak KV.

New Features

Metrics

New metrics have been added that enable you to determine the health of your Riak S2 system, as well as get reports on your storage utilization per bucket or user. The following stats items are available:

  • All calls, latencies, and counters in the S3 API
  • All calls, latencies, and counters in Stanchion
  • All Riak Erlang client operations, latencies, and counters
  • Information about the counts (active, idle, and overflow) for the process pool and connection pool
  • System information, versions, port count, and process count
  • Memory information about the riak-cs virtual machine
  • HTTP listener information: active sockets and waiting acceptors

Note: stats item names from prior to 2.0.x are not preserved; they have been renamed or removed. No backward consistency is maintained. Please see the documentation for more information.

Additional storage usage metrics are also available. . These metrics are gathered during storage calculation. Gathering these metrics is off by default, but you can turn it on by setting detailed_storage_calc to true in advanced.config. When you enable this option, you have access to information about how many manifests are writing, pending_delete, scheduled_delete and active which is not visible via the API.

Note: Metrics do not always correctly reflect actual disk usage. For instance, writing may indicate more space than is actually used. Or, for example, if an upload was cancelled in the middle, the calculation does not know how much actual storage space is consumed. In the same way, scheduled_delete also may not reflect the exact amount of disk usage because blocks might already be partially deleted by garbage collection.

riak-cs-admin

The following administration CLIs have been replaced by the riak-cs-admin command:

  • riak-cs-storage
  • riak-cs-gc
  • riak-cs-access
  • riak-cs-stanchion

The commands listed above are deprecated and will be removed in future releases.

Garbage Collection Refinements

Several new options have been added to the riak-cs-admin gc command:

  • active_delete_threshold is an option to avoid delegating manifests and block deletion to garbage collector. This option relieves garbage collector from having to delete small objects. This can optimise performance in cases where both garbage collector does not catch up with DELETE Object API calls and garbage collector’s elapsed time is dominated by small objects.[PR 1174]
  • --start and --end options have been added to the riak-cs-admin gc batch command to specify start and end in manual batch execution. Note that the --start flag on the command line will overwrite the epoch_start option in advanced.config. [PR 1147 ]
  • --leeway has been added to create a temporary leeway period whose values are used only once and not repeated at the next run, and --max-workers has been added to allow you to override the concurrency value temporarily for a single run of garbage collector. [PR 1147 ]
  • Riak S2 2.0 (and older) has a race condition where fullsync replication and garbage collection may resurrect deleted blocks without any way to delete them again. When real-time replication and replication of a garbage collection bucket entry object being dropped from the real-time queue are combined, blocks may remain on the sink side without being collected. Riak S2 2.1 introduces deterministic garbage collection to avoid fullsync replication. Additionally, garbage collection and fullsync replication run concurrently, and work on the same blocks and manifests. You can now specify the range of time using the --start and --end flags with riak-cs-admin gc batch for garbage collector in order to collect deleted objects synchronously on both sink and source sides. [PR 1147 ]
  • riak-cs-admin gc earliest-keys is available so you can find the oldest entry after epoch_start in garbage collection. With this option, you can stay informed of garbage collection progress. [PR 1160]

More information on garbage collection can be found in the documentation.

Additions

Open Source

  • A MapReduce optimisation in fetching Riak objects was introduced in Riak 2.1. Now, Riak CS 2.1 introduces an option to use that optimisation in storage calculation. It is off by default, but it can be used by setting use_2i_for_storage_calc as true in advanced.config. This reduced 50% of I/O in LevelDB. [PR 1089]
  • Erlang/OTP 17 support is now included. [PR 1245 and PR 1040]
  • A module-level hook point for limiting user access and quota usage is now available with very preliminary, simple, node-wide limiting example modules. Operators can make, plug in, or combine different modules as quota-limiting, rate-limiting or bandwidth-limiting depending on their unique requirements. [PR 1118]
  • An orphaned block scanner is now available. [PR 1133]
  • riak-cs-admin audit-bucket-ownership is a new tool to check integrity between users and buckets added. For example, it can be used in cases where a bucket is visible when listing buckets but not accessible, or a bucket is visible and exists but could not be deleted. [PR 1202]
  • The following log rotation items have been added to cuttlefish:
    • log.console.size
    • log.console.rotation
    • log.console.rotation.keep
    • log.error.rotation
    • log.error.rotation.keep
    • log.error.size

[PR 1164 and PR 97]

  • riak_cs_wm_common now has a default callback of multiple_choices, which prevents code_server from becoming a bottleneck. [PR 1181]
  • An option has been added to replace the PR=all user GET option with PR=one just before authentication. This option improves latency, especially in the presence of slow (or actually-failing) nodes blocking the whole request flow because of PR=all. When enabled, a user’s owned-bucket list is never pruned after a bucket is deleted, instead it is just marked as deleted. [PR 1191]
  • An info log has been added when starting a storage calculation batch. [PR 1238]
  • GET Bucket requests now have clearer responses. A 501 stub for Bucket lifecycle and a simple stub for Bucket requestPayment have been added. [PR 1223]
  • Several user-friendly features have been added to riak-cs-debug: fine-grained information gathering options, user-defined filtering for configuration files, and verbose output for failed commands. [PR 1236]

Enterprise

  • MDC has proxy_get, which make block objects propagate to site clusters when they are requested. Now, multibag configuration with MDC supports proxy_get. [PR 1171 and PR 25]
  • Multibag is now renamed to “Supercluster”. A bag has been a set of replicated underlying Riak clusters, which is now a member of a supercluster. riak-cs-multibag command has been renamed as riak-cs-supercluster as well. [PR 1257], [PR 1260], [PR 106], [PR 107] and [PR 31].
  • Several internal operation tools have been added to help diagnose or address issues. [PR 1145, PR 1134, and PR 1133]
  • Added a generic function for manual operations to resolve siblings of manifests and blocks, which will assist Riak Client Service Engineers with troubleshooting and solving issues. [PR 1188]

Changes

  • Dependency versions have been updated in Riak S2 and Stanchion as follows: cuttlefish 2.0.4, node_package 2.0.3, riak-erlang-client 2.1.1, lager 2.2.0, lager_syslog 2.1.1, eper 0.92 (Riak patched), cluster_info 2.0.3, riak_repl_pb_api 2.1.1, and riak_cs_multibag 2.1.0. [PR 1190, PR 1197 , PR 27, PR 1245, and PR 104].
  • Riak CS has moved from Folsom to Exometer. [PR 1165 and PR 1180]
  • Improvements have been made to error tracing for retrieving blocks from client GET requests. There is a complex logic to resolve blocks when a GET is requested from the client. First, Riak CS tries to retrieve a block with n_val=1. If it fails, a retry will be done using n_val=3. If the block cannot be resolved locally, proxy_get is enabled, and the system is configured with datacenter replication, then Riak CS will try to perform a proxied GET to the remote site. The fallback and retry logic is complex and hard to trace, especially in a faulty or unstable situation. This improvement adds error tracing for the whole sequence described above, which will help diagnose issues. Specifically, for each block, the block server stacks all errors returned from the Riak client and reports the reason for every error as well as the type of call in which the error occurred. [PR 1177]
  • Using the GET Bucket API with a specified prefix to list objects in a bucket needed optimization. It had been specifying end keys for folding objects in Riak too loosely. With this change, a tighter end key is specified for folding objects in Riak, which omits unnecessary fold in vnodes. [PR 1233]
  • A limitation to the max length of keys has been introduced. This limitation can be specified as 1024 by default, meaning no keys longer than 1024 bytes can be PUT, GET or DELETED unless max_key_length is explicitly specified as more than ‘1024’ in riak-cs.conf. If you want to preserve the old key length behaviour, you may specify the max_key_length as ‘unlimited’. [PR 1233]
  • If a faulty cluster had several nodes down, the block server misunderstood that a block was already deleted and issued a false-notfound. This could lead to block leak. The PR default has been set to ‘quorum’ in an attempt to avoid this problem. Updates have also been made to make sure at least a single replica of a block is written in one of the primary nodes by setting the PW default to ‘1’. Additionally, measures are in place to prevent the block server from crashing when “not found” errors are returned due to a particular block of an object not being found in the cluster. Instead, unreachable blocks are skipped and the remaining blocks and manifests are collected. Since the PR and PW values are increased at blocks, the availability of PUTs and through-put of garbage collection may decrease. A few Riak nodes being unreachable may prevent PUT requests from returning successfully and may prevent garbage collection from collecting all blocks until the unreachable nodes come back. [PR 1242]
  • The infinity timeout option has been set so that several functions make synchronous gen_fsm calls indefinitely, which prevents unnecessary timeouts. [PR 1249]

Bugs Fixed

  • [Issue 1097/PR 1212] When x-amz-metadata-directive=COPY was specified, Riak CS did not actually COPY the metadata of original resource. Instead, it would treat it as a REPLACE. When directed to x-amz-metadata-directive=REPLACE Content-Type, Riack CS would REPLACE it. Correct handling for the x-amz-metadata-directive has been added to PUT Object Copy API.
  • [Issue 1099/PR 1096] There was an unnecessary NextMarker in Get Bucket’s response if CommonPrefixes contained the last key. Fixed handling of uploaded parts that should be deleted after Multipart Complete Request.
  • [Issue 939/PR 1200] Copy requests without Content-Length request headers failed with 5xx errors. Such requests are now allowed without Content-Length header in Copy API calls. Additionally, Copy API calls with Content-Lengths more than zero have been given explicit errors.
  • [Issue 1143/PR 1144] Manual batch start caused the last batch time to appear to be in the future. All temporal shifts have been fixed.
  • [Issue PR 1162/PR 1163] Fix a configuration system bug where Riak CS could not start if log.syslog=on was set.
  • [Issue 1169/PR 1200] The error response of the PUT Copy API call showed the target resource path rather than the source path when the source was not found or not accessible by the request user. It now shows the source path appropriately.
  • [PR 1178] Multiple IP address descriptions under a single condition statement of a bucket policy were not being properly parsed as lists.
  • [PR 1185] If proxy_get_active was defined in riak-cs.conf as anything other than enabled or disabled, there would be excessive log output. Now, proxy_get_active also accepts non-boolean definitions.
  • [PR 1184] put_gckey_timeout was used instead of put_manifest_timeout when a delete process tried to update the status of manifests.
  • [Issue 1201/PR 1230] A single slow or silently failing node caused intermittent user fetch failure. A grace period has been added so riakc_pb_socket can attempt to reconnect.
  • [PR 1232] Warning logs were being produced for unsatisfied primary reads. Since users are objects in Riak CS and CS tries to retrieve these objects for authentication for almost every request, the retrieval option (PR=all) would fail if even one primary vnode was stopped or unresponsive and a log would be created. Given that Riak is set up to be highly available, these logs were quite noisy. Now, the “No WM route” log from prior to Riak CS 2.1 has been revived. Also, the log severity has been downgraded to debug, since it indicates a client error in all but the development phase.
  • [PR 1237] The riak-cs-admin status command exit code was non-zero, even in successful execution. It will now return zero.
  • [Issue 1097/PR 1212 and PR 4] Riak S2 did not copy the metadata of an original resource when the x-amz-metadata-directive=COPY command was used, nor when x-amz-metadata-directive was specified. Handling of the x-amz-metadata-directive command in PUT Object Copy API has been added.
  • [Issue 1097/PR 1212 and PR 4] Riak CS did not store Content-Type in COPY requests when the x-amz-metadata-directive=REPLACE command was used. Handling of the x-amz-metadata-directive command in PUT Object Copy API has been added.
  • [Issue 1097/PR 1212 and PR 4] Fixed the handling of uploaded parts that should be deleted after Multipart Complete Request.
  • [Issue 1214/PR 1246] Prior to Riak S2 2.1.0, a PUT Copy API command with identical source and destination changed user metadata (x-amz-meta-* headers) but failed to update Content-Type. Content-Type is now correctly updated by the API call.
  • [Issue PR 1261, [PR 1263] Fix riak-cs-debug to include app.config when no generated files are found when riak-cs.conf is not used.

Riak CS 2.0.1 Release Notes

General Information

This is a bugfix release.

Bug Fixes

Riak CS 2.0.0

For a complete set of release notes, upgrade instructions, and changed configuration settings, please see the Full Riak CS 2.0.0 Release Notes

General Information

  • This release updates Riak CS to work with Riak 2.0.5.
  • We have simplified the configuration system.
  • All official patches for older versions of Riak and Riak CS have been included in these releases. There is no need to apply any patches released for Riak CS 1.4.x or 1.5.x to the Riak CS 2.0.x series. Patches released for Riak CS 1.4.x or 1.5.x cannot be directly applied to Riak CS 2.0.x because the version of Erlang/OTP shipped with Riak CS has been updated in version 2.0.0.
  • Please review the complete Release Notes before upgrading.

Known Issues & Limitations

  • None.

Changes and Additions

  • Changed the name of gc_max_workers to gc.max_workers, and lowered the default value from 5 to 2 (#1110) to reduce the workload on the cs cluster.
  • Partial support of GET Location API (#1057)
  • Add very preliminary AWS v4 header authentication - without query string authentication, object chunking and payload checksum (#1064). There is still a lot of work to reliably use v4 authentication.
  • Put Enterprise deps into dependency graph (#1065)
  • Introduce Cuttlefish (#1020, #1068, #1076, #1086, #1090) (Stanchion #88, #90, #91)
  • Yessir Riak client to measure performance (#1072, #1083)
  • Inspector improvement with usage change (#1084)
  • Check signed date in S3 authentication (#1067)
  • Update cluster_info and various dependent libraries (#1087, #1088) (Stanchion #85, #87, #93)
  • Storage calculation optimization (#1089) With Riak >= 2.1 this works with use_2i_for_storage_calc flag might relieve disk read of storage calculation.

Bugfixes

  • Fix wrong webmachine log handler name (#1075)
  • Fix lager crash (#1038)
  • Fix hardcoded crashdump path (#1052)
  • Suppress unnecessary warnings (#1053)
  • Multibag simpler state transition (Multibag #21)
  • GC block deletion failure after transition to multibag environment (Multibag #19)
  • Connection closing caused errors for objects stored before the transition, after transition from single bag to multibag configuration (Multibag #18).

Deprecation Notices

  • Multi-Datacenter Replication using v2 replication support has been deprecated.
  • Old list objects which required fold_objects_for_list_keys as false have been deprecated and will be removed in the next major version.
  • Non-paginated GC in cases where gc_paginated_indexes is false has been deprecated and will be removed in the next major version.

General Notes on Upgrading to Riak CS 2.0.0

Upgrading a Riak CS system involves upgrading the underlying Riak, Riak CS and Stanchion installations. The upgrade process can be non-trivial depending on your existing system configurations and the combination of sub-system versions. This document contains general instructions and notices on upgrading the whole system to Riak CS 2.0.0.

New Configuration System

Riak 2.0.0 introduced a new configuration system (riak.conf), and as of Riak CS 2.0.0, Riak CS now supports the new configuration style. Both Riak and Riak CS still support the older style configurations through app.config and vm.args.

Riak recommends moving to the new unified configuration system, using the files riak.conf, riak-cs.conf and stanchion.conf.

Note on Legacy app.config Usage

If you choose to use the legacy app.config files for Riak CS and/or Stanchion, some parameters have changed names and must be updated.

In particular, for the Riak CS app.config:

  • cs_ip and cs_port have been combined into listener.
  • riak_ip and riak_pb_port have been combined into riak_host.
  • stanchion_ip and stanchion_port have been combined into stanchion_host.
  • admin_ip and admin_port have been combined into admin_listener.
  • webmachine_log_handler has become webmachine_access_log_handler.
  • {max_open_files, 50} has been deprecated and should be replaced with {total_leveldb_mem_percent, 30}.

For the Stanchion app.config:

  • stanchion_ip and stanchion_port have been combined into listener.
  • riak_ip and riak_port have been combined into riak_host.

Each of the above pairs follows a similar form. Where the old form used a separate IP and Port parameter, the new form combines those as {new_option, { "IP", Port}}. For example, if your legacy app.config configuration was previously:

{riak_cs, [
    {cs_ip, "127.0.0.1"},
    {cs_port, 8080 },
    . . .
]},

It should now read:

{riak_cs, [
    {listener, {"127.0.0.1", 8080}},
    . . .
]},

and so on.

Note: Upgrading from Riak CS 1.5.3 or Older

Some key objects changed names after the upgrade. Applications may need to change their behaviour due to this bugfix.

Note: Upgrading from Riak CS 1.5.0 or Older

Bucket number limitation per user have been introduced in 1.5.1. Users who have more than 100 buckets cannot create any bucket after the upgrade unless the limit is extended in the system configuration.

Note: Upgrading From Riak CS 1.4.x

An operational procedure to clean up incomplete multipart under deleted buckets is needed. Otherwise new buckets with names that used to exist can’t be created. The operation will fail with 409 Conflict.

Leeway seconds and disk space should also be carefully watched during the upgrade, because timestamp management of garbage collection was changed in the 1.5.0 release. Consult the “Leeway seconds and disk space section of 1.5 release notes for a more detailed description.

Note: Upgrading From Riak CS 1.3.x or Older

Riak supports upgrading from the two previous major versions to the latest release. Thus, this document will only cover upgrading from Riak CS versions 1.4.x and 1.5.x.

To upgrade to Riak CS 2.0.0 from versions prior to Riak CS 1.4.0, operators will need to first upgrade their system to Riak CS version 1.4.5 or 1.5.4. Upgrading to Riak CS 1.5.4 is recommended. The underlying Riak installation must also be upgraded to the Riak 1.4.x series, preferably version 1.4.12.

General Upgrade Instructions

For a complete set of release notes, upgrade instructions, and changed configuration settings, please see the Full Riak CS 2.0.0 Release Notes

All Scenarios

We recommend updating Stanchion before all other subsystems. Be careful not to have multiple live Stanchion nodes accessible from Riak CS nodes at the same time.

Repeat these steps on each node running Stanchion:

  1. Stop Stanchion
  2. Back up all Stanchion configuration files
  3. Uninstall the current Stanchion package
  4. Install the new Stanchion 2.0.0 package
  5. Migrate the Stanchion configuration (See below)
  6. Start Stanchion

Scenario: If Riak CS and Riak are both running on the same host.

Repeat these steps on every host:

  1. Stop Riak CS
  2. Stop Riak
  3. Back up all Riak and Riak CS configuration files and remove all patches
  4. Uninstall the current Riak CS package
  5. Uninstall the current Riak Riak packages
  6. Install the new Riak package
  7. Install the new Riak CS 2.0.0 package
  8. Migrate the Riak configuration (See below)
  9. Migrate the Riak CS configuration (See below)
  10. Start Riak
  11. Start Riak CS

Scenario: If Riak CS and Riak are running on separate hosts.

When Riak CS is not installed on the same host as Riak, Riak CS can be upgraded at any time while the corresponding remote Riak node is alive.

Repeat these steps on every host:

  1. Stop Riak CS
  2. Back up all configuration files and remove all patches
  3. Uninstall the current Riak CS package
  4. Install the new Riak CS 2.0.0 package
  5. Migrate the Riak CS configuration (See below)
  6. Start Riak CS

For a complete set of release notes, upgrade instructions, and changed configuration settings, please see the Full Riak CS 2.0.0 Release Notes

Riak CS 1.5.4

Fixes

  • Disable backpressure sleep riak_cs/#1041
    • Problem: When backpressure sleep is triggered due to the presence of many siblings, this can lead to even more siblings.
    • Solution: This change prevents unnecessary siblings growth in cases where (a) backpressure is triggered under high upload concurrency and (b) uploads are interleaved during backpressure sleep. This issue does not affect multipart uploads.
  • Fix an incorrect path rewrite in the S3 API caused by unnecessary URL decoding riak_cs/#1040
    • Problem: Due to the incorrect handling of URL encoding/decoding, object keys including %[0-9a-fA-F][0-9a-fA-F] (as a regular expression) or + had been mistakenly decoded. As a consequence, the former case was decoded to some other binary, while in the latter case + was replaced with a space. In both cases, there was a possibility of an implicit data overwrite. For the latter case, an overwrite occurs for an object including + in its key, e.g. foo+bar, by a different object with a name that is largely similar but replaced with space, e.g. foo bar, and vice versa.
    • Solution: Fix the incorrect handling of URL encoding/decoding. This fix also addresses riak_cs/#910 and riak_cs/#977.

Notes on Upgrading

After upgrading to Riak CS 1.5.4, objects including %[0-9a-fA-F][0-9a-fA-F] (as a regular expression) or + in their key, e.g.foo+bar, become invisible and can be seen as objects with a different name. For the former case, objects will be referred to with the unnecessarily decoded key; in the latter case, those objects will be referred to with keys in which + is replaced with a space, e.g. foo bar, by default.

The table below provides examples for URLs including %[0-9a-fA-F][0-9a-fA-F] and how they will work before and after the upgrade.

Before upgrade After upgrade
written as a%2Fkey -
read as a%2Fkey or a/key a/key
listed as a/key a/key

Examples for unique objects including + or an empty space through upgrade:

Before upgrade After upgrade
written as a+key -
read as a+key or a key a key
listed as a key a key

Examples for unique objects with an empty space in the URL:

Before upgrade After upgrade
written as a key -
read as a+key or a key a key
listed as a key a key

This fix also changes the path format in access logs from the single-URL-encoded style to the doubly-encoded URL style. Below is an example of the old style:

127.0.0.1 - - [07/Jan/2015:08:27:07 +0000] "PUT /buckets/test/objects/path1%2Fpath2%2Fte%2Bst.txt HTTP/1.1" 200 0 "" """

And here is the analogous URL in the new style:

127.0.0.1 - - [07/Jan/2015:08:27:07 +0000] "PUT /buckets/test/objects/path1%2Fpath2%2Fte%252Bst.txt HTTP/1.1" 200 0 "" ""

Note that the object path has changed from path1%2Fpath2%2Fte%2Bst.txt to path1%2Fpath2%2Fte%252Bst.txt between the two examples above.

If the old behavior is preferred, e.g. because applications using Riak CS have been written to use the older style, you can retain that behavior on upgrade by modifying your Riak CS configuration. Change the rewrite_module setting as follows:

{riak_cs, [
           %% Other settings
           {rewrite_module, riak_cs_s3_rewrite_legacy},
           %% Other settings
]}

Note: The old behavior is technically incorrect and implicitly overwrites data in the ways described above. Retain the old behavior with caution.

Riak CS 1.5.3

Changes

  • Add read_before_last_manifest_write option to help avoid sibling explosion for use cases involving high churn and concurrency on a fixed set of keys. When sibling explosion occurs, the objects stored in Riak can become very large and severely impair the functioning of the system. The trade-off in enabling this option is a latency penalty of doing an extra read before the final write of an object’s manifest to Riak. However, for use cases matching the description, the minor latency penalty is preferable to consequences of sibling explosion. riak_cs/#1011
  • Add configurable timeouts for all Riak CS interactions with Riak to provide more flexibility in operational situations. riak_cs/#1021

Fixes

  • Fix storage calculation riak_cs/#996
    • Problem: Data for deleted buckets would be included in the calculation results
    • Solution: Storage calculations no longer include deleted buckets

Known Issues

None

Download

Please see the Riak CS Downloads Page.

Feedback

We would love to hear from you. You can reach us in any of the following venues:

Riak CS 1.5.2

Changes

  • Improve logging around failures with Riak riak_cs/#987
  • Add amendment log output when storing access stats into Riak failed riak_cs/#988. This change prevents losing access stats logs in cases of temporary connection failure between Riak and Riak CS. Access logs are stored in console.log at the warning level.
  • Add script to repair invalid garbage collection manifests riak_cs/#983. There is a known issue where an active manifest would be stored in the GC bucket. This script changes invalid state to valid state.

Fixes

  • Fix Protocol Buffers connection pool (pbc_pool_master) leak riak_cs/#986.
    • Problem: Requests for non-existent buckets without an authorization header and requests for listing users make connections leak from the pool, causing the pool to eventually go empty. This bug was introduced in release 1.5.0.
    • Solution: Fix the leak by properly releasing connections.

Known Issues

None

Download

Please see the Riak CS Downloads Page

Feedback

We would love to hear from you. You can reach us in any of the following venues:

Riak CS 1.5.1

Additions

  • Bucket restrictions — Similar to S3, you can now limit the number of buckets created per user to prevent users from creating an unusually large number of buckets. More details are included here.

Changes

  • Add sleep interval after updating manifests to suppress sibling explosion riak_cs/#959. In order to suppress sibling explosion, a sleep interval is added after updating manifests. The duration of the sleep interval depends on the number of siblings. The problem is documented in more detail here.
  • Update riak-cs-debug to include information about bags in a multibag environment riak_cs/#930. Bag listing and weight information are now included in the output of the riak-cs-debug command in order to help in investigating issues in a multibag environment.
  • More efficient bucket resolution riak_cs/#951. Previously, sibling resolution logic was inefficient in cases where users had many buckets (> 1000). For the sake of optimization, resolution is now skipped entirely when no siblings are present (i.e. when there is a single value).
  • Similar to S3, add a limitation on the part number in a multipart upload riak_cs/#957. Part numbers can now range from 1 to 10,000 (inclusive).

Fixes

  • GC may stall due to riak_cs_delete_fsm deadlock riak_cs/#949
    • Problem — Garbage collection can stall when a riak_cs_delete_fsm worker process encounters a deadlock condition.
    • Solution — One of the requirements in an internal data structure was violated. This fix satisfies the requirement so that deadlock does not happen.
  • Fix wrong log directory for gathering logs on riak-cs-debug riak_cs/#953
    • Problem — Directory structure of log files gathered by riak-cs-debug was different from riak-debug.
    • Solution — The directory structure is now the same as that of riak-debug.
  • Avoid DST-aware translation from local time to GMT riak_cs/#954
    • Problem — Transformation from local time to GMT is slow, especially when performed by multiple threads. One such transformation was in the path of the GET Object API call.
    • Solution — Eliminate the transformation.
  • Use new UUID for seed of canonical ID instead of secret riak_cs/#956
    • Problem — MD5-hashed value of secret access key was used in order to generate a canonical ID, which is public information. Although MD5 reverse is not realistic, it is unnecessary and avoidable to use a secret access key for canonical ID generation.
    • Solution — Use newly generated UUID for canonical ID.
  • Set timeout as infinity to replace the default of 5000ms riak_cs/#963
    • Problem — In Riak CS 1.5.0, middleman process wrappers for Protocol Buffers sockets were introduced and call timeout to them was incorrectly set to a default of 5000 milliseconds.
    • Solution — Change the call timeout to infinity and actual timeout is controlled by Protocol Buffers processes.
  • Skip invalid state manifests in GC bucket riak_cs/#964
    • Problem — If there were active state manifests in the GC bucket the GC process crashed.
    • Solution — Skip active state manifests and make the GC process collect valid manifests.

Known Issues

None

Platforms Tested

  • Ubuntu GNU / Linux 12.04

Installation and Upgrade Notes

Per-user bucket creation restrictions

Beginning with Riak CS 1.5.1, you can limit the number of buckets that can be created per user. The default maximum number is 100. While this limitation prohibits the creation of new buckets by users, users that exceed the limit can still perform other operations, including bucket deletion. To change the default limit, add the following line to the riak_cs section of app.config:

{riak_cs, [
    %% ...
    {max_buckets_per_user, 5000},
    %% ...
]}

To avoid having a limit, set max_buckets_per_user_user to unlimited.

Download

Please see the Riak CS Downloads Page.

Feedback

We would love to hear from you. You can reach us at any of the following links:

Or via email at info@riak.com.

Riak CS 1.5.0

Additions

  • Added Multibag Technical Preview to Riak CS. More info is available here
  • A new command riak-cs-debug including cluster-info riak_cs/#769, riak_cs/#832
  • Tie up all existing commands into a new command riak-cs-admin riak_cs/#839
  • Add a command riak-cs-admin stanchion to switch Stanchion IP and port manually riak_cs/#657
  • Performance of garbage collection has been improved via Concurrent GC riak_cs/#830
  • Iterator refresh riak_cs/#805
  • fold_objects_for_list_keys made default in Riak CS riak_cs/#737, riak_cs/#785
  • Add support for Cache-Control header riak_cs/#821
  • Allow objects to be reaped sooner than leeway interval. riak_cs/#470
  • PUT Copy on both objects and upload parts riak_cs/#548
  • Update to lager 2.0.3
  • Compiles with R16B0x (Releases still by R15B01)
  • Change default value of gc_paginated_index to true riak_cs/#881
  • Add new API: Delete Multiple Objects riak_cs/#728
  • Add warning logs for manifests, siblings, bytes and history riak_cs/#915

Bugs Fixed

  • Align ERL_MAX_PORTS with Riak default: 64000 riak_cs/#636
  • Allow Riak CS admin resources to be used with OpenStack API riak_cs/#666
  • Fix path substitution code to fix Solaris source builds riak_cs/#733
  • sanity_check(true,false) logs invalid error on riakc_pb_socket error riak_cs/#683
  • Riak-CS-GC timestamp for scheduler is in the year 0043, not 2013. riak_cs/#713 fixed by riak_cs/#676
  • Excessive calls to OTP code_server process #669 fixed by riak_cs/#675
  • Return HTTP 400 if content-md5 does not match riak_cs/#596
  • /riak-cs/stats and admin_auth_enabled=false don’t work together correctly. riak_cs/#719
  • Storage calculation doesn’t handle tombstones, nor handle undefined manifest.props riak_cs/#849
  • MP initiated objects remains after delete/create buckets #475 fixed by riak_cs/#857 and stanchion/#78
  • handling empty query string on list multipart upload riak_cs/#843
  • Setting ACLs via headers at PUT Object creation riak_cs/#631
  • Improve handling of poolboy timeouts during ping requests riak_cs/#763
  • Remove unnecessary log message on anonymous access riak_cs/#876
  • Fix inconsistent ETag on objects uploaded by multipart riak_cs/#855
  • Fix policy version validation in PUT Bucket Policy riak_cs/#911
  • Fix return code of several commands, to return 0 for success riak_cs/#908
  • Fix {error, disconnected} repainted with notfound riak_cs/#929

Notes on Upgrading

Incomplete multipart uploads

riak_cs/#475 was a security issue where a newly created bucket may include unaborted or incomplete multipart uploads which was created in previous epoch of the bucket with same name. This was fixed by:

  • on creating buckets; checking if live multipart exists and if exists, return 500 failure to client.

  • on deleting buckets; trying to clean up all live multipart remains, and checking if live multipart remains (in stanchion). if exists, return 409 failure to client.

Note that a few operations are needed after upgrading from 1.4.x (or former) to 1.5.0.

  • run riak-cs-admin cleanup-orphan-multipart to cleanup all buckets. It would be safer to specify timestamp with ISO 8601 format like 2014-07-30T11:09:30.000Z as an argument. For example, in which time all CS nodes upgrade has finished. Then the cleaner does not clean up multipart uploads newer than that timestamp. Some corner cases can be prevented where multipart uploads conflicting with bucket deletion and this cleanup.

  • there might be a time period until above cleanup finished, where no client can create bucket if unfinished multipart upload remains under deleted bucket. You can find [critical] log if such bucket creation is attempted.

Leeway seconds and disk space

riak_cs/#470 changed the behaviour of object deletion and garbage collection. The timestamps in garbage collection bucket were changed from the future time when the object is to be deleted, to the current time when the object is deleted, Garbage collector was also changed to collect objects until ‘now - leeway seconds’, from collecting objects until ‘now’ previously.

Before (-1.4.x):

           t1                         t2
-----------+--------------------------+------------------->
           DELETE object:             GC triggered:
           marked as                  collects objects
           "t1+leeway"                marked as "t2"

After (1.5.0-):

           t1                         t2
-----------+--------------------------+------------------->
           DELETE object:             GC triggered:
           marked as "t1"             collects objects
           in GC bucket               marked as "t2 - leeway"

This leads that there exists a period where no objects are collected right after upgrade to 1.5.0, say, t0, until t0 + leeway . And objects deleted just before t0 won’t be collected until t0 + 2*leeway .

Also, all CS nodes which run GC should be upgraded first. CS nodes which do not run GC should be upgraded later, to let leeway second system work properly. Or stop GC while upgrading whole cluster, by running riak-cs-admin gc set-interval infinity .

Multi data center cluster should be upgraded more carefully, as to make sure GC is not running while upgrading.

Riak CS Multibag

Multibag, the ability to store object manifests and blocks in separate clusters or groups of clusters, has been added as an Enterprise feature, but it is in early preview status. proxy_get has not yet been implemented for this preview feature, so multibag is intended for a single DC only at this time.

Note: CS Multibag was renamed to CS Supercluster. More information on Supercluster Support is available here.

Known Issues and Limitations

  • If a client sends another request in the same connection while waiting for copy finish, the copy also will be aborted. This is a side effect of client disconnect detection in case of object copy. See #932 for further information.

  • Copying objects in OOS interface is not implemented.

  • Multibag is added as Enterprise feature, but it is in early preview status. proxy_get setup among clusters multibag on is not implemented yet.

Riak CS 1.4.5

Bugs Fixed

  • Fix several ‘data hiding’ bugs with the v2 list objects FSM riak_cs/788 .
  • Don’t treat HEAD requests toward BytesOut in access statistics riak_cs/791 .
  • Handle whitespace in POST/PUT XML documents riak_cs/795 .
  • Handle unicode user-names and XML riak_cs/807 .
  • Fix missing XML fields on storage usage riak_cs/808 .
  • Adjust fold-objects timeout riak_cs/811 .
  • Prune deleted buckets from user record riak_cs/812 .
  • Fix bad bucketname in storage usage riak_cs/800 .

Riak CS 1.4.4 introduced a bug (#800) where storage calculations made while running that version would have the bucket-name replaced by the string “struct”. This version fixes the bug, but can’t go back and retroactively fix the old storage calculations. Aggregations on an entire user-account should still be accurate, but you won’t be able to break-down storage by bucket, as they will all share the name “struct”.

Additions

  • Optimize the list objects v2 FSM for prefix requests riak_cs/804 .

Riak CS 1.4.4

Riak CS 1.4.4 Release Notes

Bugs Fixed

  • Create basho-patches directory riak_cs/775 .
  • sum_bucket timeout crashes all storage calculation is fixed by riak_cs/759 .
  • Failure to throttle access archiver is fixed by riak_cs/758 .
  • Access archiver crash is fixed by riak_cs/747 .

Riak CS 1.4.3

Riak CS 1.4.3 Release Notes

Bugs Fixed

  • Fix bug that reverted manifests in the scheduled_delete state to the pending_delete or active state.
  • Don’t count already deleted manifests as overwritten
  • Don’t delete current object version on overwrite with incorrect md5

Additions

  • Improve performance of manifest pruning
  • Optionally use paginated 2i for the GC daemon. This is to help prevent timeouts when collecting data that can be garbage collected.
  • Improve handling of Riak disconnects on block fetches
  • Update to Lager 2.0.1
  • Optionally prune manifests based on count, in addition to time
  • Allow multiple access archiver processes to run concurrently

Riak CS 1.4.2

Bugs Fixed

  • Fix issue with Enterprise build on Debian Linux distributions.
  • Fix source tarball build.
  • Fix access statistics bug that caused all accesses to be treated as errors.
  • Make logging in bucket listing map phase function lager version agnostic to avoid issues when using versions of Riak older than 1.4.
  • Handle undefined props field in manifests to fix issue accessing objects written with a version of Riak CS older than 1.3.0.

Additions

  • Add option to delay initial GC sweep on a node using the initial_gc_delay configuration option.
  • Append random suffix to GC bucket keys to avoid hot keys and improve performance during periods of frequent deletion.
  • Add default_proxy_cluster_id option to provide a way to specify a default cluster id to be used when the cluster id is undefined. This is to facilitate migration from the OSS version to the Enterprise version.

Riak CS 1.4.1

Bugs Fixed

  • Fix list objects crash when more than the first 1001 keys are in the pending delete state.
  • Fix crash in garbage collection daemon.
  • Fix packaging bug by updating node_package dependency.

Riak CS 1.4.0

Additions

  • Add preliminary support for the Swift API and Keystone authentication.
  • Improve performance of object listing when using Riak 1.4.0 or greater.
  • Add ability to edit user account name and email address.
  • Add support for v3 multi-data-center replication.
  • Add configurable Riak connection timeouts.
  • Add syslog support via Lager.
  • Only contact one vnode for immutable block requests.

Bugs Fixed

  • Remove unnecessary keys in GC bucket.
  • Fix query-string authentication for multi-part uploads.
  • Fix Storage Class for multi-part uploaded objects.
  • Fix etags for multi-part uploads.
  • Support reformat indexes in the Riak CS multi-backend.
  • Fix unbounded memory-growth on GET requests with a slow connection.
  • Reduce access-archiver memory use.
  • Fix 500 on object ACL HEAD request.
  • Fix semantics for concurrent upload and delete of the same key with a multipart upload.
  • Verify content-md5 header if supplied.
  • Handle transient Riak connection failures.

Riak CS 1.3.1

Bugs Fixed

  • Fix bug in handling of active object manifests in the case of overwrite or delete that could lead to old object versions being resurrected.
  • Fix improper capitalization of user metadata header names.
  • Fix issue where the S3 rewrite module omits any query parameters that are not S3 subresources. Also correct handling of query parameters so that parameter values are not URL decoded twice. This primarily affects pre-signed URLs because the access key and request signature are included as query parameters.
  • Fix for issue with init script stop.

Riak CS 1.3.0

Additions

  • Support for multipart file uploads. Parts must be in the range of 5MB-5GB.
  • Support for bucket policies using a restricted set of principals and conditions.
  • Support for returning bytes ranges of a file using the Range header.
  • Administrative commands may be segregated onto a separate interface.
  • Authentication for administrative commands may be disabled.
  • Performance and stability improvements for listing the contents of buckets.
  • Support for the prefix, delimiter, and marker options when listing the contents of a bucket.
  • Support for using Webmachine’s access logging features in conjunction with the Riak CS internal access logging mechanism.
  • Moved all administrative resources under /riak-cs.
  • Riak CS now supports packaging for FreeBSD, SmartOS, and Solaris.

Bugs Fixed

  • Fix handling of cases where buckets have siblings. Previously this resulted in 500 errors returned to the client.
  • Reduce likelihood of sibling creation when creating a bucket.
  • Return a 404 instead of a 403 when accessing a deleted object.
  • Unquote URLs to accommodate clients that URL encode / characters in URLs.
  • Deny anonymous service-level requests to avoid unnecessary error messages trying to list the buckets owned by an undefined user.

Riak CS 1.2.2

Additions

  • Full support for MDC replication

Bugs Fixed

  • Fix problem where objects with utf-8 unicode key can be neither listed nor fetched.
  • Speed up bucket_empty check and fix process leak. This bug was originally found when a user was having trouble with s3cmd rb :s3//foo --recursive. The operation first tries to delete the (potentially large) bucket, which triggers our bucket empty check. If the bucket has more than 32k items, we run out of processes unless +P is set higher (because of the leak).

Riak CS 1.2.1

Additions

  • Add reduce phase for listing bucket contents to provide backpressure when executing the MapReduce job.
  • Use prereduce during storage calculations.
  • Return 403 instead of 404 when a user attempts to list contents of nonexistent bucket.

Bugs Fixed

  • Return 403 instead of 404 when a user attempts to list contents of nonexistent bucket.
  • Do not do bucket list for HEAD or ?versioning or ?location request.

Riak CS 1.2.0

Additions

  • Add preliminary support for MDC replication
  • Quickcheck test to exercise the erlcloud library against Riak CS
  • Basic support for riak_test integration

Bugs Fixed

  • Do not expose stack traces to users on 500 errors
  • Fix issue with sibling creation on user record updates
  • Fix crash in terminate state when fsm state is not fully populated
  • Script fixes and updates in response to node_package updates

Riak CS 1.1.0

Additions

  • Update user creation to accept a JSON or XML document for user creation instead of URL encoded text string.
  • Configuration option to allow anonymous users to create accounts. In the default mode, only the administrator is allowed to create accounts.
  • Ping resource for health checks.
  • Support for user-specified metadata headers.
  • User accounts may be disabled by the administrator.
  • A new key_secret can be issued for a user by the administrator.
  • Administrator can now list all system users and optionally filter by enabled or disabled account status.
  • Garbage collection for deleted and overwritten objects.
  • Separate connection pool for object listings with a default of 5 connections.
  • Improved performance for listing all objects in a bucket.
  • Statistics collection and querying.
  • DTrace probing.

Bugs Fixed

  • Check for timeout when checking out a connection from poolboy.
  • PUT object now returns 200 instead of 204.
  • Fixes for Dialyzer errors and warnings.
  • Return readable error message with 500 errors instead of large webmachine backtraces.

Riak CS 1.0.2

Additions

Riak CS 1.0.1

Bugs Fixed

  • Default content-type is not passed into function to handle PUT request body
  • Requests hang when a node in the Riak cluster is unavailable
  • Correct inappropriate use of riak_cs_utils:get_user by riak_moss_acl_utils:get_owner_data

Riak CS 1.0.0

Additions

  • Subsystem for calculating user access and storage usage
  • Fixed-size connection pool of Riak connections
  • Use a single Riak connection per request to avoid deadlock conditions
  • Object ACLs
  • Management for multiple versions of a file manifest
  • Configurable block size and max content length
  • Support specifying non-default ACL at bucket creation time

Bugs Fixed

  • Fix PUTs for zero-byte files
  • Fix fsm initialization race conditions
  • Canonicalize the entire path if there is no host header, but there are tokens
  • Fix process and socket leaks in get fsm

Riak CS 0.1.2

Bugs Fixed

  • Return 403 instead of 503 for invalid anonymous or signed requests.
  • Properly clean up processes and connections on object requests.

Riak CS 0.1.1

Bugs Fixed

  • HEAD requests always result in a 403 Forbidden.
  • s3cmd info on a bucket object results in an error due to missing ACL document.
  • Incorrect atom specified in riak_moss_wm_utils:parse_auth_header.
  • Bad match condition used in riak_moss_acl:has_permission/2.

Riak CS 0.1.0

Additions

  • Bucket-level access control lists
  • User records have been modified so that an system-wide unique email address is required to create a user.
  • User creation requests are serialized through stanchion to be certain the email address is unique.
  • Bucket creation and deletion requests are serialized through stanchion to ensure bucket names are unique in the system.
  • The stanchion serialization service is now required to be installed and running for the system to be fully operational.
  • The concept of an administrative user has been added to the system. The credentials of the administrative user must be added to the app.config files for moss and stanchion.
  • User credentials are now created using a url-safe base64 encoding module.

Bugs Fixed

  • s3cmd info fails due to missing last-modified key in return document.
  • s3cmd get of 0 byte file fails.
  • Bucket creation fails with status code 415 using the AWS Java SDK.

Known Issues

  • Object-level access control lists have not yet been implemented.

Riak CS 0.0.3

Additions

  • Support for the s3cmd subcommands sync, du, and rb
  • Return valid size and checksum for each object when listing bucket objects.
  • Changes so that a bucket may be deleted if it is empty.
  • Changes so a subdirectory path can be specified when storing or retrieving files.
  • Make buckets private by default
  • Support the prefix query parameter
  • Enhance process dependencies for improved failure handling

Bugs Fixed

  • URL decode keys on put so they are represented correctly. This eliminates confusion when objects with spaces in their names are listed and when attempting to access them.
  • Properly handle zero-byte files
  • Reap all processes during file puts

Known Issues

  • Buckets are marked as /private/ by default, but globally unique bucket names are not enforced. This means that two users may create the same bucket and this could result in unauthorized access and unintentional overwriting of files. This will be addressed in a future release by ensuring that bucket names are unique across the system.