Improvements to garbage collector (fix #39) #135

Merged
lx merged 4 commits from cleanups into main 2021-11-09 13:03:35 +00:00
3 changed files with 247 additions and 147 deletions
Showing only changes of commit 08b1e8a7ea - Show all commits

View file

@ -29,6 +29,7 @@
- [Design](./design/index.md)
- [Related Work](./design/related_work.md)
- [Internals](./design/internals.md)
- [Design draft](./design/design_draft.md)
- [Development](./development/index.md)
- [Setup your environment](./development/devenv.md)

View file

@ -0,0 +1,162 @@
# Design draft
**WARNING: this documentation is a design draft which was written before Garage's actual implementation.
The general principle are similar, but details have not been updated.**
#### Modules
- `membership/`: configuration, membership management (gossip of node's presence and status), ring generation --> what about Serf (used by Consul/Nomad) : https://www.serf.io/? Seems a huge library with many features so maybe overkill/hard to integrate
- `metadata/`: metadata management
- `blocks/`: block management, writing, GC and rebalancing
- `internal/`: server to server communication (HTTP server and client that reuses connections, TLS if we want, etc)
- `api/`: S3 API
- `web/`: web management interface
#### Metadata tables
**Objects:**
- *Hash key:* Bucket name (string)
- *Sort key:* Object key (string)
- *Sort key:* Version timestamp (int)
- *Sort key:* Version UUID (string)
- Complete: bool
- Inline: bool, true for objects < threshold (say 1024)
- Object size (int)
- Mime type (string)
- Data for inlined objects (blob)
- Hash of first block otherwise (string)
*Having only a hash key on the bucket name will lead to storing all file entries of this table for a specific bucket on a single node. At the same time, it is the only way I see to rapidly being able to list all bucket entries...*
**Blocks:**
- *Hash key:* Version UUID (string)
- *Sort key:* Offset of block in total file (int)
- Hash of data block (string)
A version is defined by the existence of at least one entry in the blocks table for a certain version UUID.
We must keep the following invariant: if a version exists in the blocks table, it has to be referenced in the objects table.
We explicitly manage concurrent versions of an object: the version timestamp and version UUID columns are index columns, thus we may have several concurrent versions of an object.
Important: before deleting an older version from the objects table, we must make sure that we did a successfull delete of the blocks of that version from the blocks table.
Thus, the workflow for reading an object is as follows:
1. Check permissions (LDAP)
2. Read entry in object table. If data is inline, we have its data, stop here.
-> if several versions, take newest one and launch deletion of old ones in background
3. Read first block from cluster. If size <= 1 block, stop here.
4. Simultaneously with previous step, if size > 1 block: query the Blocks table for the IDs of the next blocks
5. Read subsequent blocks from cluster
Workflow for PUT:
1. Check write permission (LDAP)
2. Select a new version UUID
3. Write a preliminary entry for the new version in the objects table with complete = false
4. Send blocks to cluster and write entries in the blocks table
5. Update the version with complete = true and all of the accurate information (size, etc)
6. Return success to the user
7. Launch a background job to check and delete older versions
Workflow for DELETE:
1. Check write permission (LDAP)
2. Get current version (or versions) in object table
3. Do the deletion of those versions NOT IN A BACKGROUND JOB THIS TIME
4. Return succes to the user if we were able to delete blocks from the blocks table and entries from the object table
To delete a version:
1. List the blocks from Cassandra
2. For each block, delete it from cluster. Don't care if some deletions fail, we can do GC.
3. Delete all of the blocks from the blocks table
4. Finally, delete the version from the objects table
Known issue: if someone is reading from a version that we want to delete and the object is big, the read might be interrupted. I think it is ok to leave it like this, we just cut the connection if data disappears during a read.
("Soit P un problème, on s'en fout est une solution à ce problème")
#### Block storage on disk
**Blocks themselves:**
- file path = /blobs/(first 3 hex digits of hash)/(rest of hash)
**Reverse index for GC & other block-level metadata:**
- file path = /meta/(first 3 hex digits of hash)/(rest of hash)
- map block hash -> set of version UUIDs where it is referenced
Usefull metadata:
- list of versions that reference this block in the Casandra table, so that we can do GC by checking in Cassandra that the lines still exist
- list of other nodes that we know have acknowledged a write of this block, usefull in the rebalancing algorithm
Write strategy: have a single thread that does all write IO so that it is serialized (or have several threads that manage independent parts of the hash space). When writing a blob, write it to a temporary file, close, then rename so that a concurrent read gets a consistent result (either not found or found with whole content).
Read strategy: the only read operation is get(hash) that returns either the data or not found (can do a corruption check as well and return corrupted state if it is the case). Can be done concurrently with writes.
**Internal API:**
- get(block hash) -> ok+data/not found/corrupted
- put(block hash & data, version uuid + offset) -> ok/error
- put with no data(block hash, version uuid + offset) -> ok/not found plz send data/error
- delete(block hash, version uuid + offset) -> ok/error
GC: when last ref is deleted, delete block.
Long GC procedure: check in Cassandra that version UUIDs still exist and references this block.
Rebalancing: takes as argument the list of newly added nodes.
- List all blocks that we have. For each block:
- If it hits a newly introduced node, send it to them.
Use put with no data first to check if it has to be sent to them already or not.
Use a random listing order to avoid race conditions (they do no harm but we might have two nodes sending the same thing at the same time thus wasting time).
- If it doesn't hit us anymore, delete it and its reference list.
Only one balancing can be running at a same time. It can be restarted at the beginning with new parameters.
#### Membership management
Two sets of nodes:
- set of nodes from which a ping was recently received, with status: number of stored blocks, request counters, error counters, GC%, rebalancing%
(eviction from this set after say 30 seconds without ping)
- set of nodes that are part of the system, explicitly modified by the operator using the web UI (persisted to disk),
is a CRDT using a version number for the value of the whole set
Thus, three states for nodes:
- healthy: in both sets
- missing: not pingable but part of desired cluster
- unused/draining: currently present but not part of the desired cluster, empty = if contains nothing, draining = if still contains some blocks
Membership messages between nodes:
- ping with current state + hash of current membership info -> reply with same info
- send&get back membership info (the ids of nodes that are in the two sets): used when no local membership change in a long time and membership info hash discrepancy detected with first message (passive membership fixing with full CRDT gossip)
- inform of newly pingable node(s) -> no result, when receive new info repeat to all (reliable broadcast)
- inform of operator membership change -> no result, when receive new info repeat to all (reliable broadcast)
Ring: generated from the desired set of nodes, however when doing read/writes on the ring, skip nodes that are known to be not pingable.
The tokens are generated in a deterministic fashion from node IDs (hash of node id + token number from 1 to K).
Number K of tokens per node: decided by the operator & stored in the operator's list of nodes CRDT. Default value proposal: with node status information also broadcast disk total size and free space, and propose a default number of tokens equal to 80%Free space / 10Gb. (this is all user interface)
#### Constants
- Block size: around 1MB ? --> Exoscale use 16MB chunks
- Number of tokens in the hash ring: one every 10Gb of allocated storage
- Threshold for storing data directly in Cassandra objects table: 1kb bytes (maybe up to 4kb?)
- Ping timeout (time after which a node is registered as unresponsive/missing): 30 seconds
- Ping interval: 10 seconds
- ??
#### Links
- CDC: <https://www.usenix.org/system/files/conference/atc16/atc16-paper-xia.pdf>
- Erasure coding: <http://web.eecs.utk.edu/~jplank/plank/papers/CS-08-627.html>
- [Openstack Storage Concepts](https://docs.openstack.org/arch-design/design-storage/design-storage-concepts.html)
- [RADOS](https://ceph.com/wp-content/uploads/2016/08/weil-rados-pdsw07.pdf)

View file

@ -1,158 +1,95 @@
**WARNING: this documentation is more a "design draft", which was written before Garage's actual implementation. The general principle is similar but details have not yet been updated.**
# Internals
#### Modules
## Overview
- `membership/`: configuration, membership management (gossip of node's presence and status), ring generation --> what about Serf (used by Consul/Nomad) : https://www.serf.io/? Seems a huge library with many features so maybe overkill/hard to integrate
- `metadata/`: metadata management
- `blocks/`: block management, writing, GC and rebalancing
- `internal/`: server to server communication (HTTP server and client that reuses connections, TLS if we want, etc)
- `api/`: S3 API
- `web/`: web management interface
TODO: write this section
#### Metadata tables
- The Dynamo ring
**Objects:**
- CRDTs
- *Hash key:* Bucket name (string)
- *Sort key:* Object key (string)
- *Sort key:* Version timestamp (int)
- *Sort key:* Version UUID (string)
- Complete: bool
- Inline: bool, true for objects < threshold (say 1024)
- Object size (int)
- Mime type (string)
- Data for inlined objects (blob)
- Hash of first block otherwise (string)
- Consistency model of Garage tables
*Having only a hash key on the bucket name will lead to storing all file entries of this table for a specific bucket on a single node. At the same time, it is the only way I see to rapidly being able to list all bucket entries...*
**Blocks:**
- *Hash key:* Version UUID (string)
- *Sort key:* Offset of block in total file (int)
- Hash of data block (string)
A version is defined by the existence of at least one entry in the blocks table for a certain version UUID.
We must keep the following invariant: if a version exists in the blocks table, it has to be referenced in the objects table.
We explicitly manage concurrent versions of an object: the version timestamp and version UUID columns are index columns, thus we may have several concurrent versions of an object.
Important: before deleting an older version from the objects table, we must make sure that we did a successfull delete of the blocks of that version from the blocks table.
Thus, the workflow for reading an object is as follows:
1. Check permissions (LDAP)
2. Read entry in object table. If data is inline, we have its data, stop here.
-> if several versions, take newest one and launch deletion of old ones in background
3. Read first block from cluster. If size <= 1 block, stop here.
4. Simultaneously with previous step, if size > 1 block: query the Blocks table for the IDs of the next blocks
5. Read subsequent blocks from cluster
Workflow for PUT:
1. Check write permission (LDAP)
2. Select a new version UUID
3. Write a preliminary entry for the new version in the objects table with complete = false
4. Send blocks to cluster and write entries in the blocks table
5. Update the version with complete = true and all of the accurate information (size, etc)
6. Return success to the user
7. Launch a background job to check and delete older versions
Workflow for DELETE:
1. Check write permission (LDAP)
2. Get current version (or versions) in object table
3. Do the deletion of those versions NOT IN A BACKGROUND JOB THIS TIME
4. Return succes to the user if we were able to delete blocks from the blocks table and entries from the object table
To delete a version:
1. List the blocks from Cassandra
2. For each block, delete it from cluster. Don't care if some deletions fail, we can do GC.
3. Delete all of the blocks from the blocks table
4. Finally, delete the version from the objects table
Known issue: if someone is reading from a version that we want to delete and the object is big, the read might be interrupted. I think it is ok to leave it like this, we just cut the connection if data disappears during a read.
("Soit P un problème, on s'en fout est une solution à ce problème")
#### Block storage on disk
**Blocks themselves:**
- file path = /blobs/(first 3 hex digits of hash)/(rest of hash)
**Reverse index for GC & other block-level metadata:**
- file path = /meta/(first 3 hex digits of hash)/(rest of hash)
- map block hash -> set of version UUIDs where it is referenced
Usefull metadata:
- list of versions that reference this block in the Casandra table, so that we can do GC by checking in Cassandra that the lines still exist
- list of other nodes that we know have acknowledged a write of this block, usefull in the rebalancing algorithm
Write strategy: have a single thread that does all write IO so that it is serialized (or have several threads that manage independent parts of the hash space). When writing a blob, write it to a temporary file, close, then rename so that a concurrent read gets a consistent result (either not found or found with whole content).
Read strategy: the only read operation is get(hash) that returns either the data or not found (can do a corruption check as well and return corrupted state if it is the case). Can be done concurrently with writes.
**Internal API:**
- get(block hash) -> ok+data/not found/corrupted
- put(block hash & data, version uuid + offset) -> ok/error
- put with no data(block hash, version uuid + offset) -> ok/not found plz send data/error
- delete(block hash, version uuid + offset) -> ok/error
GC: when last ref is deleted, delete block.
Long GC procedure: check in Cassandra that version UUIDs still exist and references this block.
Rebalancing: takes as argument the list of newly added nodes.
- List all blocks that we have. For each block:
- If it hits a newly introduced node, send it to them.
Use put with no data first to check if it has to be sent to them already or not.
Use a random listing order to avoid race conditions (they do no harm but we might have two nodes sending the same thing at the same time thus wasting time).
- If it doesn't hit us anymore, delete it and its reference list.
Only one balancing can be running at a same time. It can be restarted at the beginning with new parameters.
#### Membership management
Two sets of nodes:
- set of nodes from which a ping was recently received, with status: number of stored blocks, request counters, error counters, GC%, rebalancing%
(eviction from this set after say 30 seconds without ping)
- set of nodes that are part of the system, explicitly modified by the operator using the web UI (persisted to disk),
is a CRDT using a version number for the value of the whole set
Thus, three states for nodes:
- healthy: in both sets
- missing: not pingable but part of desired cluster
- unused/draining: currently present but not part of the desired cluster, empty = if contains nothing, draining = if still contains some blocks
Membership messages between nodes:
- ping with current state + hash of current membership info -> reply with same info
- send&get back membership info (the ids of nodes that are in the two sets): used when no local membership change in a long time and membership info hash discrepancy detected with first message (passive membership fixing with full CRDT gossip)
- inform of newly pingable node(s) -> no result, when receive new info repeat to all (reliable broadcast)
- inform of operator membership change -> no result, when receive new info repeat to all (reliable broadcast)
Ring: generated from the desired set of nodes, however when doing read/writes on the ring, skip nodes that are known to be not pingable.
The tokens are generated in a deterministic fashion from node IDs (hash of node id + token number from 1 to K).
Number K of tokens per node: decided by the operator & stored in the operator's list of nodes CRDT. Default value proposal: with node status information also broadcast disk total size and free space, and propose a default number of tokens equal to 80%Free space / 10Gb. (this is all user interface)
See this presentation (in French) for some first information:
<https://git.deuxfleurs.fr/Deuxfleurs/garage/src/branch/main/doc/talks/2020-12-02_wide-team/talk.pdf>
#### Constants
## Garbage collection
- Block size: around 1MB ? --> Exoscale use 16MB chunks
- Number of tokens in the hash ring: one every 10Gb of allocated storage
- Threshold for storing data directly in Cassandra objects table: 1kb bytes (maybe up to 4kb?)
- Ping timeout (time after which a node is registered as unresponsive/missing): 30 seconds
- Ping interval: 10 seconds
- ??
A faulty garbage collection procedure has been the cause of
[critical bug #39](https://git.deuxfleurs.fr/Deuxfleurs/garage/issues/39).
This precise bug was fixed in the code, however there are potentially more
general issues with the garbage collector being too eager and deleting things
too early. This has been the subject of
[PR #135](https://git.deuxfleurs.fr/Deuxfleurs/garage/pulls/135).
This section summarizes the discussions on this topic.
#### Links
Rationale: we want to ensure Garage's safety by making sure things don't get
deleted from disk if they are still needed. Two aspects are involved in this.
### 1. Garbage collection of table entries (in `meta/` directory)
The `Entry` trait used for table entries (defined in `tables/schema.rs`)
defines a function `is_tombstone()` that returns `true` if that entry
represents an entry that is deleted in the table. CRDT semantics by default
keep all tombstones, because they are necessary for reconciliation: if node A
has a tombstone that supersedes a value `x`, and node B has value `x`, A has to
keep the tombstone in memory so that the value `x` can be properly deleted at
node `B`. Otherwise, due to the CRDT reconciliation rule, the value `x` from B
would flow back to A and a deleted item would reappear in the system.
Here, we have some control on the nodes involved in storing Garage data.
Therefore we have a garbage collector that is able to delete tombstones UNDER
CERTAIN CONDITIONS. This garbage collector is implemented in `table/gc.rs`. To
delete a tombstone, the following condition has to be met:
- All nodes responsible for storing this entry are aware of the existence of
the tombstone, i.e. they cannot hold another version of the entry that is
superseeded by the tombstone. This ensures that deleting the tombstone is
safe and that no deleted value will come back in the system.
Garage makes use of Sled's atomic operations (such as compare-and-swap and
transactions) to ensure that only tombstones that have been correctly
propagated to other nodes are ever deleted from the local entry tree.
This GC is safe in the following sense: no non-tombstone data is ever deleted
from Garage tables.
**However**, there is an issue with the way this interacts with data
rebalancing in the case when a partition is moving between nodes. If a node has
some data of a partition for which it is not responsible, it has to offload it.
However that offload process takes some time. In that interval, the GC does not
check with that node if it has the tombstone before deleting the tombstone, so
perhaps it doesn't have it and when the offload finally happens, old data comes
back in the system.
**PR 135 mostly fixes this** by implementing a 24-hour delay before anything is
garbage collected in a table. This works under the assumption that rebalances
that follow data shuffling terminate in less than 24 hours.
**However**, in distributed systems, it is generally considered a bad practice
to make assumptions that information propagates in a certain time interval:
this consists in making a synchrony assumption, meaning that we are basically
assuming a computing model that has much stronger properties than otherwise. To
maximize the applicability of Garage, we would like to remove this assumption,
and implement a system where time does not play a role. To do this, we would
need to find a way to safely disable the GC when data is being shuffled around,
and safely detect that the shuffling has terminated and thus the GC can be
resumed. This introduces some complexity to the protocol and hasn't been
tackled yet.
### 2. Garbage collection of data blocks (in `data/` directory)
Blocks in the data directory are reference-counted. In Garage versions before
PR #135, blocks could get deleted from local disk as soon as their reference
counter reached zero. We had a mechanism to not trigger this immediately at the
rc-reaches-zero event, but the cleanup could be triggered by other means (for
example by a block repair operation...). PR #135 added a safety measure so that
blocks never get deleted in a 10 minute interval following the time when the RC
reaches zero. This is a measure to make impossible race conditions such as #39.
We would have liked to use a larger delay (e.g. 24 hours), but in the case of a
rebalance of data, this would have led to the disk utilization to explode
during the rebalancing, only to shrink again after 24 hours. The 10-minute
delay is a compromise that gives good security while not having this problem of
disk space explosion on rebalance.
- CDC: <https://www.usenix.org/system/files/conference/atc16/atc16-paper-xia.pdf>
- Erasure coding: <http://web.eecs.utk.edu/~jplank/plank/papers/CS-08-627.html>
- [Openstack Storage Concepts](https://docs.openstack.org/arch-design/design-storage/design-storage-concepts.html)
- [RADOS](https://ceph.com/wp-content/uploads/2016/08/weil-rados-pdsw07.pdf)