`.era` files and Req/Resp protocols use framed formats - aligning the
database with these makes for less recompression work overall as gossip
is sent only once while req/resp repeats (potentially) - this also
allows efficient pruning-to-era where snappy-recompression is the major
cycle thief.
* era: load blocks and states
Era files contain finalized history and can be thought of as an
alternative source for block and state data that allows clients to avoid
syncing this information from the P2P network - the P2P network is then
used to "top up" the client with the most recent data. They can be
freely shared in the community via whatever means (http, torrent, etc)
and serve as a permanent cold store of consensus data (and, after the
merge, execution data) for history buffs and bean counters alike.
This PR gently introduces support for loading blocks and states in two
cases: block requests from rest/p2p and frontfilling when doing
checkpoint sync.
The era files are used as a secondary source if the information is not
found in the database - compared to the database, there are a few key
differences:
* the database stores the block indexed by block root while the era file
indexes by slot - the former is used only in rest, while the latter is
used both by p2p and rest.
* when loading blocks from era files, the root is no longer trivially
available - if it is needed, it must either be computed (slow) or cached
(messy) - the good news is that for p2p requests, it is not needed
* in era files, "framed" snappy encoding is used while in the database
we store unframed snappy - for p2p2 requests, the latter requires
recompression while the former could avoid it
* front-filling is the process of using era files to replace backfilling
- in theory this front-filling could happen from any block and
front-fills with gaps could also be entertained, but our backfilling
algorithm cannot take advantage of this because there's no (simple) way
to tell it to "skip" a range.
* front-filling, as implemented, is a bit slow (10s to load mainnet): we
load the full BeaconState for every era to grab the roots of the blocks
- it would be better to partially load the state - as such, it would
also be good to be able to partially decompress snappy blobs
* lookups from REST via root are served by first looking up a block
summary in the database, then using the slot to load the block data from
the era file - however, there needs to be an option to create the
summary table from era files to fully support historical queries
To test this, `ncli_db` has an era file exporter: the files it creates
should be placed in an `era` folder next to `db` in the data directory.
What's interesting in particular about this setup is that `db` remains
as the source of truth for security purposes - it stores the latest
synced head root which in turn determines where a node "starts" its
consensus participation - the era directory however can be freely shared
between nodes / people without any (significant) security implications,
assuming the era files are consistent / not broken.
There's lots of future improvements to be had:
* we can drop the in-memory `BlockRef` index almost entirely - at this
point, resident memory usage of Nimbus should drop to a cool 500-600 mb
* we could serve era files via REST trivially: this would drop backfill
times to whatever time it takes to download the files - unlike the
current implementation that downloads block by block, downloading an era
at a time almost entirely cuts out request overhead
* we can "reasonably" recreate detailed state history from almost any
point in time, turning an O(slot) process into O(1) effectively - we'll
still need caches and indices to do this with sufficient efficiency for
the rest api, but at least it cuts the whole process down to minutes
instead of hours, for arbitrary points in time
* CI: ignore failures with Nim-1.6 (temporary)
* test fixes
Co-authored-by: Ștefan Talpalaru <stefantalpalaru@yahoo.com>
One more step on the journey to reduce `BlockRef` usage across the
codebase - this one gets rid of `StateData` whose job was to keep track
of which block was last assigned to a state - these duties have now been
taken over by `latest_block_root`, a fairly recent addition that
computes this block root from state data (at a small cost that should be
insignificant)
99% mechanical change.
When performing trusted node sync, historical access is limited to
states after the checkpoint.
Reindexing restores full historical access by replaying historical
blocks against the state and storing snapshots in the database.
The process can be initiated or resumed at any point in time.
This PR names and documents the concept of the archive: a range of slots
for which we have degraded functionality in terms of historical access -
in particular:
* we don't support rewinding to states in this range
* we don't keep an in-memory representation of the block dag
The archive de-facto exists in a trusted-node-synced node, but this PR
gives it a name and drops the in-memory digest index.
In order to satisfy `GetBlocksByRange` requests, we ensure that we have
blocks for the entire archive period via backfill. Future versions may
relax this further, adding a "pre-archive" period that is fully pruned.
During by-slot searches in the archive (both for libp2p and rest
requests), an extra database lookup is used to covert the given `slot`
to a `root` - future versions will avoid this using era files which
natively are indexed by `slot`. That said, the lookup is quite
fast compared to the actual block loading given how trivial the table
is - it's hard to measure, even.
A collateral benefit of this PR is that checkpoint-synced nodes will see
100-200MB memory usage savings, thanks to the dropped in-memory cache -
future pruning work will bring this benefit to full nodes as well.
* document chaindag storage architecture and assumptions
* look up parent using block id instead of full block in clearance
(future-proofing the code against a future in which blocks come from era
files)
* simplify finalized block init, always writing the backfill portion to
db at startup (to ensure lookups work as expected)
* preallocate some extra memory for finalized blocks, to avoid immediate
realloc
To calculate the deltas correctly, the `process_inactivity_updates` function
must be called before the rewards and penalties processing code in order to
update the `inactivity_scores` field in the state. This would have required
duplicating more logic from the spec in the ncli modules, so I've decided to
pay the price of introducing a run-time copy of the state at each epoch which
eliminates the need to duplicate logic (both for this fix and the previous one).
Other changes:
* Fixes for the read-only mode of the `BeaconChainDb`
* Fix an uint64 underflow in the debug output procedure for printing
balance deltas
* Allow Bellatrix states in the reward computation helpers
Streamline lookup with Forky and BeaconBlockFork (then we can do the
same for era)
We use type to avoid conditionals, as fork is often already known at a
"higher" level.
* load blockid before loading block by root - this is needed to map root
to slot and will eventually be done via block summary table for "old"
blocks
Co-authored-by: tersec <tersec@users.noreply.github.com>
Update several `ncli_db` commands to run in readOnly mode, allowing them
to be used with a running instance - in particular era export.
* export all eras by default
* skip already-exported eras
* clean up / document init
* drop `immutable_validators` data (pre-altair)
* document versions where data is first added
* avoid needlessly loading genesis block data on startup
* add a few more internal database consistency checks
* remove duplicate state root lookup on state load
* comment
* Store finalized block roots in database (3s startup)
When the chain has finalized a checkpoint, the history from that point
onwards becomes linear - this is exploited in `.era` files to allow
constant-time by-slot lookups.
In the database, we can do the same by storing finalized block roots in
a simple sparse table indexed by slot, bringing the two representations
closer to each other in terms of conceptual layout and performance.
Doing so has a number of interesting effects:
* mainnet startup time is improved 3-5x (3s on my laptop)
* the _first_ startup might take slightly longer as the new index is
being built - ~10s on the same laptop
* we no longer rely on the beacon block summaries to load the full dag -
this is a lot faster because we no longer have to look up each block by
parent root
* a collateral benefit is that we no longer need to load the full
summaries table into memory - we get the RSS benefits of #3164 without
the CPU hit.
Other random stuff:
* simplify forky block generics
* fix withManyWrites multiple evaluation
* fix validator key cache not being updated properly in chaindag
read-only mode
* drop pre-altair summaries from `kvstore`
* recreate missing summaries from altair+ blocks as well (in case
database has lost some to an involuntary restart)
* print database startup timings in chaindag load log
* avoid allocating superfluos state at startup
* use a recursive sql query to load the summaries of the unfinalized
blocks
This is an alternative take on https://github.com/status-im/nimbus-eth2/pull/3107
that aims for more minimal interventions in the spec modules at the expense of
duplicating more of the spec logic in ncli_db.
Overhaul of era files, including documentation and reference
implementations
* store blocks, then state, then slot indices for easy lookup at low
cost
* document era file rationale
* altair+ support in era writer
Introduced in #3171, it turns out we can just follow the block headers
to achieve the same effect
* leaves the constant in the code so as to avoid confusion when reading
database that had the constant written (such as the fleet nodes and
other unstable users)
In the ChainDAG, 3 block pointers are kept: genesis, tail and head. This
PR adds one more block pointer: the backfill block which represents the
block that has been backfilled so far.
When doing a checkpoint sync, a random block is given as starting point
- this is the tail block, and we require that the tail block has a
corresponding state.
When backfilling, we end up with blocks without corresponding states,
hence we cannot use `tail` as a backfill pointer - there is no state.
Nonetheless, we need to keep track of where we are in the backfill
process between restarts, such that we can answer GetBeaconBlocksByRange
requests.
This PR adds the basic support for backfill handling - it needs to be
integrated with backfill sync, and the REST API needs to be adjusted to
take advantage of the new backfilled blocks when responding to certain
requests.
Future work will also enable moving the tail in either direction:
* pruning means moving the tail forward in time and removing states
* backwards means recreating past states from genesis, such that
intermediate states are recreated step by step all the way to the tail -
at that point, tail, genesis and backfill will match up.
* backfilling is done when backfill != genesis - later, this will be the
WSS checkpoint instead
Turning uncompressed pubkeys into cooked ones is fast, but unnecessary -
this should avoid a little work for every signature validation we do by
pre-loading them at startup.
* ncli_db: add putState, putBlock
These tools allow modifying an existing nimbus database for the purpose
of recovery or reorg, moving the head, tail and genesis to arbitrary
points.
* remove potentially expensive `putState` in `BeaconStateDB`
* introduce `latest_block_root` which computes the root of the latest
applied block from the `latest_block_header` field (instead of passing
it in separately)
* avoid some unnecessary BeaconState copies during init
* discover https://github.com/nim-lang/Nim/issues/19094
* prefer `HashedBeaconState` in a few places to avoid recomputing state
root
* fetch latest block root from state when creating blocks
* harden `get_beacon_proposer_index` against invalid slots and document
* move random spec function tests to `test_spec.nim`
* avoid unnecessary state root computation before block proposal
* fix stack overflow crash in REST/debug/getStateV2
* introduce `ForkyXxx` for generic type matching of `Xxx` across
branches (SomeHashedBeaconState -> ForkyHashedBeaconState et al) -
`Some` is already used for other types of type classes
* consolidate function naming in BeaconChainDB, use some generics
* import `forks.nim` from other spec modules and move `Forked*` helpers
around to resolve circular imports
* remove `ForkedBeaconState`, use `ForkedHashedBeaconState` throughout
(less data shuffling between the types)
* fix several cases of states being stored on stack in tests, causing
random failures on some platforms
* remove reading json support from ncli - this should be ported to the
rest json reading instead (doesn't currently work because stack sizes)
* reorganize ssz dependencies
This PR continues the work in
https://github.com/status-im/nimbus-eth2/pull/2646,
https://github.com/status-im/nimbus-eth2/pull/2779 as well as past
issues with serialization and type, to disentangle SSZ from eth2 and at
the same time simplify imports and exports with a structured approach.
The principal idea here is that when a library wants to introduce SSZ
support, they do so via 3 files:
* `ssz_codecs` which imports and reexports `codecs` - this covers the
basic byte conversions and ensures no overloads get lost
* `xxx_merkleization` imports and exports `merkleization` to specialize
and get access to `hash_tree_root` and friends
* `xxx_ssz_serialization` imports and exports `ssz_serialization` to
specialize ssz for a specific library
Those that need to interact with SSZ always import the `xxx_` versions
of the modules and never `ssz` itself so as to keep imports simple and
safe.
This is similar to how the REST / JSON-RPC serializers are structured in
that someone wanting to serialize spec types to REST-JSON will import
`eth2_rest_serialization` and nothing else.
* split up ssz into a core library that is independendent of eth2 types
* rename `bytes_reader` to `codec` to highlight that it contains coding
and decoding of bytes and native ssz types
* remove tricky List init overload that causes compile issues
* get rid of top-level ssz import
* reenable merkleization tests
* move some "standard" json serializers to spec
* remove `ValidatorIndex` serialization for now
* remove test_ssz_merkleization
* add tests for over/underlong byte sequences
* fix broken seq[byte] test - seq[byte] is not an SSZ type
There are a few things this PR doesn't solve:
* like #2646 this PR is weak on how to handle root and other
dontSerialize fields that "sometimes" should be computed - the same
problem appears in REST / JSON-RPC etc
* Fix a build problem on macOS
* Another way to fix the macOS builds
Co-authored-by: Zahary Karadjov <zahary@gmail.com>
The spec imports are a mess to work with, so this branch cleans them up
a bit to ensure that we avoid generic sandwitches and that importing
stuff generally becomes easier.
* reexport crypto/digest/presets because these are part of the public
symbol set of the rest of the spec types
* don't export `merge` types from `base` - this causes circular deps
* fix circular deps in `ssz/spec_types` - this is the first step in
disentangling ssz from spec
* be explicit about phase0 vs altair - longer term, `altair` will become
the "natural" type set, then merge and so on, so no point in giving
`phase0` special preferential treatment
* update validator key cache on startup
Versions prior to 1.1.0 do not write a validator key cache at all.
Versions from 1.4.0 and upwards require an immutable validator key cache
to verify blocks - normally, block verification fills the cache but that
assumes that at least one block was verified by a version that has the
key cache.
Taken together, this breaks direct upgrades from anything <1.1.0 to
1.4.0.
The fix is simply to refresh fill the cache from an existing state on
startup.
* also log serious block validation failures at info level
* Implement split preset/config support
This is the initial bulk refactor to introduce runtime config values in
a number of places, somewhat replacing the existing mechanism of loading
network metadata.
It still needs more work, this is the initial refactor that introduces
runtime configuration in some of the places that need it.
The PR changes the way presets and constants work, to match the spec. In
particular, a "preset" now refers to the compile-time configuration
while a "cfg" or "RuntimeConfig" is the dynamic part.
A single binary can support either mainnet or minimal, but not both.
Support for other presets has been removed completely (can be readded,
in case there's need).
There's a number of outstanding tasks:
* `SECONDS_PER_SLOT` still needs fixing
* loading custom runtime configs needs redoing
* checking constants against YAML file
* yeerongpilly support
`build/nimbus_beacon_node --network=yeerongpilly --discv5:no --log-level=DEBUG`
* load fork epoch from config
* fix fork digest sent in status
* nicer error string for request failures
* fix tools
* one more
* fixup
* fixup
* fixup
* use "standard" network definition folder in local testnet
Files are loaded from their standard locations, including genesis etc,
to conform to the format used in the `eth2-networks` repo.
* fix launch scripts, allow unknown config values
* fix base config of rest test
* cleanups
* bundle mainnet config using common loader
* fix spec links and names
* only include supported preset in binary
* drop yeerongpilly, add altair-devnet-0, support boot_enr.yaml
* add blockchain_dag altair database reading; add rollback tests; fix some unnecessary type conversions
* remove debugging scaffolding
* proposeSignedBlock() will need to be async for merge; introduce altair types to VC
* introduce immutable Altair BeaconState
* add database support for Altair blocks and states
* add tests for Altair get/put/contains/delete state
* enable blockchain_dag Altair state database storing
* properly return error on getting missing altair block
* use ForkedHashedBeaconState in StateData
* fix FAR_FUTURE_EPOCH -> slot overflow; almost always use assign()
* avoid stack allocation in maybeUpgradeStateToAltair()
* create and use dispatch functions for check_attester_slashing(), check_proposer_slashing(), and check_voluntary_exit()
* use getStateRoot() instead of various state.data.hbsPhase0.root
* remove withStateVars.hashedState(), which doesn't work as a design anymore
* introduce spec/datatypes/altair into beacon_chain_db
* fix inefficient codegen for getStateField(largeStateField)
* state_transition_slots() doesn't either need/use blocks or runtime presets
* combine process_slots(HBS)/state_transition_slots(HBS) which differ only in last-slot htr optimization
* getStateField(StateData, ...) was replaced by getStateField(ForkedHashedBeaconState, ...)
* fix rollback
* switch some state_transition(), process_slots, makeTestBlocks(), etc to use ForkedHashedBeaconState
* remove state_transition(phase0.HashedBeaconState)
* remove process_slots(phase0.HashedBeaconState)
* remove state_transition_block(phase0.HashedBeaconState)
* remove unused callWithBS(); separate case expression from if statement
* switch back from nested-ref-object construction to (ref Foo)(Bar())
* write uncompressed validator keys to database
Loading 150k+ validator keys on startup in compressed format takes a lot
of time - better store them in uncompressed format which makes behaviour
just after startup faster / more predictable.
* refactor cached validator key access
* fix isomorphic cast to work with non-var instances
* remove cooked pubkey cache - directly use database cache in chaindag
as well (one less cache to keep in sync)
* bump blscurve, introduce loadValid for known-to-be-valid keys
* don't consider legacy database when writing state - this read is slow
on kvstore
* avoid epoch transition when there's an exact match in cache already
* simplify init to only consider checkpoint states
This reverts commit eebc828778.
Adding a separate file turns out not to be enough. This PR reverts the
separate file change.
Another theory is that the large kvstore table causes cache thrashing -
all database connections share a common page cache which would explain
the poor performance of the separate file solution.
The V1 table structure shows great improvements in performance, but if
there's an old `kvstore` without rowid:s, these benefits are nullified:
reorgs during writes and deletes remain expensive (even if the
degradation is reduced somewhat).
This PR creates the tables in a new file instead, and uses the old file
as a read-only store - this has several interesting properties:
* the old database is left completely untouched - this guarantees that
downgrades work smooth (they'll only need to resync their missing
portions)
* starting sync after this PR means only a v1 database is created
* v0 databases stick around - no migration is performed (for now)
Future PR:s can introduce migration of the data from one database to
another - a simply copy will take hours which is downtime we want to
avoid - at that point, it might make sense to migrate straight to era
files instead.
* Revert "Revert "Upgrade database schema" (#2570)"
This reverts commit 6057c2ffb4.
* ssz: fix loading empty lists into existing instances
Not a problem earlier because we didn't reuse instances
* bump nim-eth
* bump nim-web3
The `kvstore` design we're using now turns out to not be the best way to
use `sqlite` - in particular, there are some significant benefits to
using rowid in certain situations and to keep data in separate tables.
With this branch, there are massive improvements in startup time
(seconds instead of minutes) and state/block storage and pruning times
(milliseconds instead of seconds) - these improvements can in particular
be seen on slow drives and translate directly into better attestation
performance.
* update kvstore to new keyspace design
* remove `DirStoreRef` and the hidden `--state-db-kind` option - this
was an experiment to store large blobs in files, but with the new
kvstore, there's no compelling reason to do so
* remove `DbMap` - unused and would need updating for new keyspace
design
* introduce separate tables for each data type (blocks, states etc)
* remove "WITHOUT ROWID" pessimization for tables with large blobs
* close DbSeq statements explicitly (and earlier)
* store beacon block summaries in separate table, without SSZ
compression and load them all with single query on startup
* stop storing backwards compat full states
* mark genesis beacon block as trusted
* avoid faststreams when loading SSZ data
* remove `DisagreementBehavior` (unused)
This patch writes a full genesis state to `kvstore` if one was missing,
which fixes 1.2.0 restarting sync when upgrading from 1.1.0, or when
downgrading to a pre-1.1.0 release.
* Reset cached indices when resetting cache on SSZ read
When deserializing into an existing structure, the cache should be
cleared - goes for json also. Also improve error messages.
* initial immutable validator database factoring
* remove changes from chain_dag: this abstraction properly belongs in beacon_chain_db
* add merging mutable/immutable validator portions; individually test database roundtripping of immutable validators and states-sans-immutable-validators
* update test summaries
* use stew/assign2 instead of Nim assignment
* add reading/writing of immutable validators in chaindag
* remove unused import
* replace chunked k/v store of immutable validators with per-row SQL table storage
* use List instead of HashList
* un-stub some ncli_db code so that it uses
* switch HashArray to array; move BeaconStateNoImmutableValidators from datatypes to beacon_chain_db
* begin only-mutable-part state storage
* uncomment some assigns
* work around https://github.com/nim-lang/Nim/issues/17253
* fix most of the issues/oversights; local sim runs again
* fix test suite by adding missing beaconstate field to copy function
* have ncli bench also store immutable validators
* extract some immutable-validator-specific code from the beacon chain db module
* add more rigorous database state roundtripping, with changing validator sets
* adjust ncli_db to use new schema
* simplify putState/getState by moving all immutable validator accounting into beacon state DB
* remove redundant test case and move code to immutable-beacon-chain module
* more efficient, but still brute-force, mutable+immutable validator merging
* reuse BeaconState in getState
* ensure HashList/HashArray caches are cleared when reusing getState buffers; add ncli_db and a unit test to verify this
* HashList.clear() -> HashList.clearCache()
* only copy incrementally necessary immutable validators
* increase strictness of test cases and fix/work around resulting HashList cache invalidation issues
* remove explanatory scaffolding
* allow for storage of full (with all validators) states for backwards/forwards-compatibility
* adjust DbSeq type usage
* store full, with-validators, state every 64 epochs to enable reverting versions
* reduce memory allocation and intermediate objects in state storage codepath
* eliminate allocation/copying through intermediate BeaconStateNoImmutableValidators objects
* skip benchmarking initial genesis-validator-heavy state store
* always store new-style state and sometimes old-style state
* document intent behind BeaconState/Validator type-punnery
* more accurate failure message on SQLite in-memory database initialization failure
* database state storage benchmarking via ncli_db
* more cleanups from immutable validator state branch
* unexport some eth2_network constants and remove unused variables/templates
* make two PeerScore constants public
* checkpoint database at end of each slot
To avoid spending time on synchronizing with the file system while doing
processing, the manual checkpointing mode turns off fsync during
processing and instead checkpoints the database when the slot has ended.
From an sqlite perspecitve, in WAL mode this guaranees database
consistency but may lead to data loss which is fine - anything missing
from the beacon chain database can be recovered on the next startup.
* log sync status and delay in slot start message
* bump
* Handle some web3 timeouts better
* Add support for developer .env files
* Eth1 improvements; Mainnet genesis state
Notable changes:
* The deposits table have been removed from the database. The client
will no longer process all deposits on start-up.
* The network metadata now includes a "state snapshot" of the deposit
contract. This allows the client to skip syncing deposits made prior
to the snapshot (i.e. genesis). Suitable metadata added for Pyrmont
and Mainnet.
* The Eth1 monitor won't be started unless there are validators attached
to the node.
* The genesis detection code is now optional and disabled by default
* Bugfix: The client should not produce blocks that will fail validation
when it hasn't downloaded the latest deposits yet
* Bugfix: Work around the database corruption affecting Pyrmont nodes
* Remove metadata for Toledo and Medalla
This introcudes a cache for block summaries, useful for instantiating
the block dag on startup, bringing medalla startup times down from
minutes to seconds.
This is something of a temporary band-aid that would be obsoleted by a
finalized block store.
* Concentrate all sensitive writeFile/createPath calls in one place.
Fix eth2_network_simulation for Windows.
* Remove artifacts.
* fix import
Co-authored-by: Jacek Sieka <jacek@status.im>
* fix invalid state root being written to database
When rewinding state data, the wrong block reference would be used when
saving the state root - this would cause state loading to fail by
loading a different state than expected, preventing blocks to be
applied.
* refactor state loading and saving to consistently use and set
StateData block
* avoid rollback when state is missing from database (as opposed to
being partially overwritten and therefore in need of rollback)
* don't store state roots for empty slots - previously, these were used
as a cache to avoid recalculating them in state transition, but this has
been superceded by hash tree root caching
* don't attempt loading states / state roots for non-epoch slots, these
are not saved to the database
* simplify rewinder and clean up funcitions after caches have been
reworked
* fix chaindag logscope
* add database reload metric
* re-enable clearance epoch tests
* names
hash_tree_root was turning up when running beacon_node, turns out to be
repeated hash_tree_root invocations - this pr brings them back down to
normal.
this PR caches the root of a block in the SignedBeaconBlock object -
this has the potential downside that even invalid blocks will be hashed
(as part of deserialization) - later, one could imagine delaying this
until checks have passed
there's also some cleanup of the `cat=` logs which were applied randomly
and haphazardly, and to a large degree are duplicated by other
information in the log statements - in particular, topics fulfill the
same role
* cleanups
* fix ncli state root check flag
* add block dump to ncli_db
* limit ncli_db benchmark length
* tone down finalization logs
* introduce trusted blocks
We only store blocks whose signature we've verified in the database - as
such, there's no need to check it again, and most importantly, no need
to deserialize the signature when loading from database.
50x startup time improvement, 200x block load time improvement.
* fix rewinding when deposits have invalid signature
* speed up ancestor iteration by avoiding copy
* avoid deserializing signatures for trusted data
* load blocks lazily when rewinding (less memory used)
* chronicles workarounds
* document trustedbeaconblock
* 3gb vs 12gb for 4000 epochs of witti
* 3-4x sync blocks/sec performance improvement on my FDE SSD drive
* generate less quirky code for primitive types
* reorder ssz
* split into hash_trees and ssz_serialization, roughly, for hashing and
IO
* move bitseqs into ssz (from stew)
* clean up imports
* docs, imports
Jacek Sieka
zah
tersec
Mamy Ratsimbazafy
Eugene Kabanov